hypermnesia-mcp

---

Your coding agent forgets you every time you close the session. Every architecture decision you explained. Every debugging session where you traced a bug through four layers of abstraction. Every "remember, we decided to use event sourcing, not CRUD" correction. Gone. Next session, your agent is a stranger to its own tools.

Cortex is a cross-platform persistent memory engine for AI coding agents, built on computational neuroscience. Codex, Gemini CLI, Claude Code, and any local stdio MCP host can use the same remember/recall, knowledge-graph, consolidation, and wiki tools. Claude Code's plugin adds automatic capture and injection hooks; other hosts use the same memory through explicit tool calls.

It runs entirely on your machine — a local SQLite database by default (zero setup, no services to install), or PostgreSQL + pgvector when you want it. A 22 MB embedding model, no LLM in the retrieval loop, no data leaving localhost.

36 neuroscience mechanisms · 52 memory tools · 9 lifecycle hooks · a self-curating, continuously-groomed per-project wiki — all local, all open-source.

---

Getting Started

Cortex ships as a single-click MCP bundle (.mcpb). Download the latest hypermnesia-mcp.mcpb from Releases, then open it in Claude Desktop — Settings → Extensions installs it in one click.

It runs immediately on the built-in SQLite backend: zero configuration, no database to provision, nothing to set up. Memory persists to a local file under ~/.claude/methodology/. That's the whole install.

Claude Cowork works the same zero-setup way: the sandboxed environment is detected automatically (CLAUDE_ENVIRONMENT=cowork) and Cortex uses the local SQLite store — no PostgreSQL required.

Want PostgreSQL + pgvector instead (for very large stores or a shared team database)? It's a single configuration field — see Configuration below. SQLite is the default; PostgreSQL is opt-in.

Claude Code plugin (marketplace):

claude plugin marketplace add cdeust/Cortex
claude plugin install hypermnesia-mcp

Upgrading from the cortex plugin? The plugin was renamed hypermnesia-mcp in v4.15.0 (a community-directory name collision with an unrelated cortex plugin): claude plugin uninstall cortex && claude plugin install hypermnesia-mcp — your memories and configuration are untouched, storage paths do not change.

That is the whole install — zero configuration, no PostgreSQL, no system packages. The postInstall provisions Python dependencies and selects the local SQLite store (~/.claude/methodology/memory.db); the store schema auto-creates on first use. The embedding model is not downloaded at install time — it fetches lazily on first use (~100 MB, one-time; see PRIVACY.md) and runs fully offline afterwards. The plugin path registers all lifecycle hooks (session-start context injection, per-prompt auto-recall, auto-capture, compaction checkpointing, the autonomous wiki cycle) and the /cortex-setup-project command.

An existing PostgreSQL install is never downgraded: the installer detects a configured DATABASE_URL, a prior PostgreSQL backend marker, or a reachable local cortex database and keeps using it across plugin updates.

Upgrading the plugin to PostgreSQL (optional):

bash <plugin-dir>/scripts/install-plugin.sh --postgres   # <plugin-dir> = the installed plugin root

In one line: PostgreSQL adds connection-pooled concurrency (two psycopg_pool latency classes), server-side PL/pgSQL WRRF fusion, and pgvector HNSW ANN indexing — worth it for very large stores or a shared team database (see Under the Hood); the memory tools and retrieval contract are identical on both backends. After upgrading, run /cortex-setup-project once — it handles pgvector setup, database creation, the embedding-model pre-cache, profile building, codebase seeding, and hook registration.

What SQLite mode does not do (honest disclosure): the WRRF fusion runs in-process instead of server-side, without HNSW ANN indexing (fine at personal-store scale, slower at very large scale); and three PostgreSQL-only hook enrichments degrade to silent no-ops — cross-agent team-decision injection (agent_briefing, plus the banner's Team Decisions section), file-based preemptive context (preemptive_context), and pipeline symbol heat-bumps (pipeline_impact_bump). Session-start banners, auto-recall injection, auto-capture, checkpoints, and all 52 memory tools work on both backends.

Verify any install:

python3 -m mcp_server.doctor

The check list is backend-aware: on SQLite it verifies Python, the store opens, a writable methodology dir, and the pool-capacity invariant; on PostgreSQL it additionally checks the PG driver, DATABASE_URL, connection, and extensions. Exit 0 means ready.

Clone + setup script:

git clone https://github.com/cdeust/Cortex.git && cd Cortex
bash scripts/setup.sh        # macOS / Linux
python3 scripts/setup.py     # Windows / cross-platform

Docker:

git clone https://github.com/cdeust/Cortex.git && cd Cortex
docker build -t cortex-runtime -f docker/Dockerfile .
docker run -it \
  -v $(pwd):/workspace \
  -v cortex-pgdata:/var/lib/postgresql/17/data \
  -v ~/.claude:/home/cortex/.claude-host:ro \
  cortex-runtime

PyPI (uvx / pip) — best-effort hook-free compatibility:

uvx hypermnesia-mcp          # run the MCP server directly
pip install hypermnesia-mcp  # or install into your environment

The server is published on PyPI as hypermnesia-mcp (registry name io.github.cdeust/hypermnesia-mcp). Claude Code remains the primary integration because it adds the lifecycle hooks; PyPI is the best-effort hook-free stdio compatibility channel for Gemini CLI, Codex CLI, and other MCP hosts.

WSL / TLS client-cert / remote PostgreSQL: See deployment scenarios.

---

Use with other MCP hosts

The MCP server is host-agnostic: any host that can launch a stdio process gets the full tool surface — remember, recall, the wiki, navigation, consolidation, all 52 tools — on the default local SQLite store. What is not portable are the 9 lifecycle hooks, which are Claude Code plugin machinery. The server itself does not import or require those hooks at startup.

What works where (honest matrix):

Capability	Claude Code plugin	Local stdio hosts (Gemini CLI, Codex CLI, ChatGPT desktop, Cursor, Windsurf, VS Code, Agents SDK)	ChatGPT web
All 52 memory tools (remember, recall, wiki, navigation, consolidation, triggers, rules)	✅	✅	❌ — Cortex does not ship a remote HTTPS endpoint
SQLite default store / PostgreSQL opt-in	✅	✅	❌ — a remote deployment and per-user storage/auth model would be required
Auto-capture of significant tool output	✅ (PostToolUse hook)	❌ — store explicitly with remember	❌
Session-start context injection	✅ (SessionStart hook)	❌ — call recall yourself	❌
Per-prompt auto-recall	✅	❌	❌
Compaction checkpoints	✅	❌	❌
Autonomous wiki cycle (headless worker)	✅	❌ — run consolidate / curate_wiki manually	❌
Cognitive profiling (query_methodology)	✅	⚠️ profiles are mined from Claude Code session logs under ~/.claude/; without them the profile is empty	❌

In one sentence: on Claude Code memory is ambient (hooks capture and inject automatically); on every other host memory is manual-tool-driven — the agent stores and retrieves when instructed, and nothing happens between prompts.

The launch command on every host is the PyPI package (the [sqlite] extra enables sqlite-vec vector search; without it the store still works, with vector search disabled):

uvx --from "hypermnesia-mcp[sqlite]" hypermnesia-mcp

Gemini CLI — this repo ships a gemini-extension.json:

gemini extensions install https://github.com/cdeust/Cortex

Or add it to ~/.gemini/settings.json directly:

{
  "mcpServers": {
    "cortex": {
      "command": "uvx",
      "args": ["--from", "hypermnesia-mcp[sqlite]", "hypermnesia-mcp"]
    }
  }
}

OpenAI Codex and ChatGPT desktop — native local plugin (recommended). The repository now carries an isolated Codex marketplace and an exact 10-tool lean MCP surface. Claude Code remains the primary integration and retains its automatic hooks, custom agent, full tool profile, and unchanged marketplace. Pre-install the same published package once so the plugin's first uvx handshake can reuse the local uv cache instead of spending its startup budget downloading the Python environment. The bundled server also declares a 180-second startup ceiling, backed by a 110.46-second clean-cache launch measured on 2026-08-02 on local macOS 26.5.1 arm64 with uv 0.8.19 (the clean ubuntu-latest CI run completed in 23.87 seconds):

uv tool install "hypermnesia-mcp[sqlite]"
codex plugin marketplace add cdeust/Cortex
codex plugin add hypermnesia-mcp-codex@cortex-codex-plugins

Restart the ChatGPT desktop app and start a new task after installation. See the Codex plugin guide for the host boundary and the public-directory requirements Cortex deliberately does not claim.

Direct Codex MCP configuration (fallback). After the same pre-installation, register the executable directly:

uv tool install "hypermnesia-mcp[sqlite]"
codex mcp add cortex --env CORTEX_MEMORY_STORE_BACKEND=sqlite -- hypermnesia-mcp

codex mcp add registers the executable and SQLite backend. For production use, extend that generated entry with explicit startup and tool timeouts. Codex CLI, the Codex IDE extension, and the ChatGPT desktop app's local Codex host share ~/.codex/config.toml; the recommended complete entry is:

[mcp_servers.cortex]
command = "hypermnesia-mcp"
startup_timeout_sec = 30
tool_timeout_sec = 600

[mcp_servers.cortex.env]
CORTEX_MEMORY_STORE_BACKEND = "sqlite"

Verify discovery with codex mcp list and then /mcp inside a Codex session. ChatGPT web is a different surface: it does not read local Codex configuration and accepts MCP tools through hosted plugins backed by remote Streamable HTTP servers. Cortex deliberately does not claim that deployment model today; exposing a local personal-memory database through a remote endpoint would require an explicit authentication, tenancy, and privacy design.

The server preserves FastMCP's diagnostic startup banner (including the FASTMCP_SHOW_SERVER_BANNER setting) but disables its network update probe: an MCP stdio handshake must succeed offline and must not fail because of proxy-specific HTTP extras. This does not freeze FastMCP indefinitely. Upgrade the installed tool explicitly; the new Cortex release and its declared dependency set are then resolved together:

uv tool upgrade hypermnesia-mcp

Repository installs remain reproducible through uv.lock, whose dependency updates are reviewed through normal pull requests.

Validation scope: CI runs the installed production console entry point through the MCP lifecycle under representative Claude, Gemini, and Codex client identities for both the full and exact 10-tool lean profiles, then separately invokes pinned vendor CLIs to parse the Claude plugin, Gemini extension, and recommended Codex configuration. This proves the protocol and configuration contracts; it does not simulate an authenticated model turn inside each vendor UI.

Cursor — .cursor/mcp.json (project) or ~/.cursor/mcp.json (global) — and Windsurf — ~/.codeium/windsurf/mcp_config.json — take the same mcpServers block as Gemini above.

VS Code — .vscode/mcp.json:

{
  "servers": {
    "cortex": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "hypermnesia-mcp[sqlite]", "hypermnesia-mcp"]
    }
  }
}

OpenAI Agents SDK (Python):

from agents.mcp import MCPServerStdio

async with MCPServerStdio(
    name="cortex",
    params={
        "command": "uvx",
        "args": ["--from", "hypermnesia-mcp[sqlite]", "hypermnesia-mcp"],
    },
) as server:
    agent = Agent(name="Assistant", mcp_servers=[server])

uvx requires uv; pip install "hypermnesia-mcp[sqlite]" + the hypermnesia-mcp console script works identically. GUI hosts that don't inherit your shell PATH may need the absolute path from which uvx.

---

Configuration

Cortex needs no configuration to run — the SQLite backend is the default and requires nothing. Two optional settings let you change the storage backend; in the single-click bundle they appear as fields in Claude Desktop's extension settings, and everywhere else they map to environment variables.

Setting	Env var	Default	What it does
Storage backend	CORTEX_MEMORY_STORE_BACKEND	sqlite*	sqlite runs fully local with zero setup. postgresql uses an external PostgreSQL + pgvector database (set the URL below). auto tries PostgreSQL and falls back to SQLite.
PostgreSQL URL	CORTEX_MEMORY_DATABASE_URL	(empty)	Only used when the backend is postgresql or auto. Example: postgresql://user:password@host:5432/cortex. Leave empty to stay on SQLite. Treated as sensitive.

* The single-click bundle pins the backend to sqlite through the manifest. If you run the server directly (clone / Docker) without setting the variable, the underlying code default is auto — it tries PostgreSQL and falls back to SQLite.

That's the entire surface most users touch. Both backends expose the same 52 memory tools (55 with the optional automatised-pipeline + prd-spec-generator integrations) and the same retrieval contract; PostgreSQL adds server-side PL/pgSQL fusion and HNSW indexing that pays off at very large scale. Every other knob uses the CORTEX_MEMORY_ prefix — see mcp_server/infrastructure/memory_config.py.

---

Examples

A live, end-to-end run on the SQLite backend — store three memories, recall them by meaning, then check the store. The output is taken from the in-process FastMCP client (recall lists trimmed to the top hit). The harness writes with force: true for determinism, and the demo store already held a few earlier memories — so memory_stats totals exceed the three inserted here.

1 — Store a memory. It is stored with a heat score (force: true skips the dedup write-gate to keep the demo deterministic; omit it and a near-duplicate would be gated).

remember({
  content: "Cortex stores memory in a local SQLite database by default — zero setup, no PostgreSQL required.",
  tags: ["architecture", "decision"],
  force: true
})
// → { stored: true, memory_id: 490, action: "stored", heat: 0.796 }

2 — Recall by meaning, not keywords. The fused retrieval ranks the relevant memory first.

recall({ query: "how does cortex store memory by default?" })
// → memories[0] = {
//     content: "Cortex stores memory in a local SQLite database by default — zero setup, no PostgreSQL required.",
//     score: 0.0167, heat: 0.846, tags: ["architecture", "decision"]
//   }

3 — A different query surfaces a different memory. Stored "Anchored memories survive context compaction with maximum priority."; this recall puts it on top.

recall({ query: "what survives context compaction?" })
// → memories[0] = {
//     content: "Anchored memories survive context compaction with maximum priority.",
//     heat: 0.565, tags: ["compaction"]
//   }

4 — Inspect the store. has_vector_search: true confirms semantic search is live on SQLite.

memory_stats({})
// → { total_memories: 14, episodic_count: 8, semantic_count: 6,
//     avg_heat: 0.942, has_vector_search: true }

You rarely call these by hand: the lifecycle hooks (plugin install) inject the right memories at session start and capture new ones as you work. The tools are there when you want explicit control — anchor to pin an architecture constraint, consolidate to run a maintenance cycle, narrative to get the project's story so far.

---

What's new

v4.13.0 — grooming becomes continuous instead of session-bound. A measured 76-day wiki-citation and lesson-promotion silence (invisible until this release could measure it) is closed by wiring a recurring citation-reconciliation pass and a lesson-promotion backlog count into every consolidation cycle, a new READ_ONLY get_grooming_health tool (backlog + staleness per grooming type, on demand), a one-line session_start nudge past a sourced staleness threshold, and an opt-in scheduled scripts/groomer.py entry point (dry-run by default) for anyone who wants the maintenance pass on its own cron/launchd cadence. Also fixed: headless wiki writes now go through the same governed path as interactive writes, and — found on the way — a CRITICAL latent bug where the anchor-page frontmatter template's status: living violated the wiki.pages CHECK constraint since its origin, silently failing the database sync of every interactively-authored curate_wiki page in production. 49 memory tools (52 with upstream integrations). See CHANGELOG.md for the full write-up.

v4.0.0 — the neuroscience model complete (13 new mechanisms). Cortex fills the remaining cognitive-science gaps so memory spans encoding → consolidation → retrieval → forgetting with a grounded mechanism at every stage: source/reality monitoring (C1) with a confabulation gate, recollection-vs-familiarity dual-process retrieval (C2), claim-conflict monitoring (A2), goal maintenance (A1), attentional-salience gating, habituation (E1), fear-extinction inhibitory learning (E2), stress/arousal encoding-gain modulation, a predictive-coding forward model, value/reward-weighted retention, procedural (skill) memory (B1), two-phase NREM/REM sleep consolidation (F1), and cued targeted reactivation (F2). One new MCP tool (recall_skills); each mechanism is cited to published work and exposed as a live system vital. 44 memory tools (47 with upstream integrations).

v3.23.0 — single-click bundle + registry-indexer build fix. Cortex now ships as an MCP bundle (.mcpb) with a uv runtime and a selectable storage backend — SQLite by default, PostgreSQL optional — so it installs in one click with zero setup. Also: a neuro-cortex-memory console script so uv run neuro-cortex-memory resolves from a checkout; registry indexers that build with uv sync and launch via that script now start the server and register all standalone tools without a PostgreSQL connection (so tools/list answers inside a DB-less container).

v3.21.0 — visualization extracted to cortex-viz. The entire visualization stack — the galaxy graph, execution trace, the Knowledge / Board / Wiki / Pipeline views, and their HTTP server — moves to a standalone companion MCP, cortex-viz, which reads this same store read-only. Cortex is a focused memory engine again (−50k lines). Breaking: the open_visualization, get_methodology_graph, and query_workflow_graph MCP tools are removed from Cortex — install cortex-viz to get them back (its /cortex-visualize skill replaces the old one). 46 MCP tools remain; no memory, retrieval, or wiki behaviour changed; full suite green (3214 tests).

→ Full changelog and release notes

---

The science under the hood

Cortex doesn't store memories the way a database stores rows. It treats them the way a brain treats experiences. Every mechanism traces to a published paper — a 97-reference bibliography (docs/papers/bibliography.md).

Memories have temperature. Every memory starts hot. Access it and it stays hot; ignore it and it cools. Below a threshold it compresses: full text → summary → keywords → fades entirely. This is rate-distortion optimal forgetting — the framework your brain uses to decide what's worth keeping. Important memories resist compression; surprising ones get a heat boost; boring, redundant ones quietly disappear. (Anderson & Lebiere 1998; Ebbinghaus 1885)

Storage has a gatekeeper. Not everything deserves to be remembered. Cortex maintains a predictive model of what it already knows and only stores information that violates its expectations. Tell it the same thing twice and the write gate blocks the second attempt. This is predictive coding — the mechanism your neocortex uses to filter sensory input. Only prediction errors get through. (Friston 2005; Bastos et al. 2012)

Retrieval changes the memory. When you recall a memory in a new context, Cortex compares the retrieval context against the storage context, and if there's enough mismatch it reconsolidates — updates the memory to reflect what's true now. Nader et al. showed in 2000 that retrieved memories become labile and can be rewritten. Your codebase evolves, and so do Cortex's memories of it. (Dudai 2012; Nader et al. 2000)

Emotional memories are stronger. Frustration during debugging, urgency in a production incident — Cortex detects emotional valence and encodes those memories with more force. They decay slower, compress later, and surface faster, like how you remember your worst outage in vivid detail but not last Tuesday's standup. (Qasim et al. 2023; Hebb 1955)

Background consolidation runs like sleep. When you're away, a consolidation cycle decays old memories, compresses verbose ones, promotes recurring patterns into general knowledge (episodic → semantic transfer), discovers entity relationships, and runs "dream replay" where related memories are compared and new connections emerge. (McClelland et al. 1995; Foster & Wilson 2006; Buzsáki 2015)

Similar memories stay distinct. Pattern separation, modeled on the dentate gyrus, keeps "Tuesday's standup" separate from "Wednesday's standup" even though they're nearly identical — without it, retrieval returns the same generic match for every similar query. (Leutgeb et al. 2007; Yassa & Stark 2011)

Forgetting is active, not just passive decay. Beyond the slow cooling of the heat model, Cortex runs two independent dopaminergic forgetting circuits from Drosophila: a permanent one that erodes a memory's trace under chronic interference (Rac1), and a transient one that blocks retrieval of a memory without erasing it (DAMB). Interference-heavy, low-value memories are actively removed rather than left to linger — the same way your brain prunes what competes without paying off. (Davis & Zhong 2017; Sabandal et al. 2021)

The two arXiv-ready papers go deeper: Thermodynamic Memory vs. Flat-Importance Stores (PDF, 34 pages) · Stage-Aware Context Assembly (PDF, 39 pages).

---

What this actually feels like

Monday. You spend an hour debugging a webhook handler. After tracing through four layers, you find the root cause: a race condition in the Redis session store where TTL expiry can fire between the auth check and the permission lookup. You discuss the fix with Claude, decide on an approach, implement it. Session ends.

Thursday. Different project, but a user reports intermittent logouts. You open Claude. Before you even describe the bug, Cortex has already injected three memories: Monday's race-condition analysis, a decision from two weeks ago to use Redis for all session state, and a lesson from an older session about TTL edge cases in distributed caches.

Claude doesn't just have your conversation history. It has context — it connects the current problem to past decisions and skips the part where you re-explain your architecture.

Three weeks later. Those debugging sessions have consolidated into a general pattern: "authentication edge cases involving TTL-based caches." The specific Redis commands compressed to a summary, the debugging steps faded, the principle survived. Your next auth issue starts with institutional knowledge, not a blank page.

---

Retrieval that actually works

We tested Cortex against three published benchmarks. All scores are retrieval-only — no LLM reader in the evaluation loop. We measure whether the right memory shows up, not whether a model can generate a good answer from it.

LongMemEval — can you find a fact from 40 sessions ago?

LongMemEval (Wu et al., ICLR 2025): 500 human-curated questions embedded in ~40 sessions of conversation history (~115k tokens). The paper's best retrieval hit 78.4% Recall@10.

	Cortex	What it means
Recall@10	98.2%	The right memory shows up in the top 10 for nearly every question
MRR	0.915	The correct memory is usually ranked first or second — retrieval rank only, no LLM reader

n=500, E1 v3 verification campaign — per-row JSONs with code SHAs in benchmarks/results/ablation/longmemeval-s_v3/. Re-verified on a clean DB 2026-06-10.

Category	MRR	R@10
Single-session (assistant)	1.000	100.0%
Multi-session reasoning	0.962	100.0%
Knowledge updates	0.925	100.0%
Temporal reasoning	0.926	98.5%
Single-session (user)	0.814	94.3%
Single-session (preference)	0.668	93.3%

Knowledge updates score near-perfect because the retrieval stack's recency signal and update-intent routing push the newest version of a fact above older ones.

LoCoMo — trick questions and multi-hop reasoning

LoCoMo (Maharana et al., ACL 2024): 1,986 questions across 10 conversations — adversarial trick questions, multi-hop queries needing evidence from multiple turns, and temporal reasoning.

	Cortex	What it means
Recall@10	91.5%	Right memory in top 10 over 9 times out of 10
MRR	0.805	The correct memory is typically ranked first — retrieval rank only, no LLM reader

n=1986, BASELINE_NO_CONSOLIDATION, post-plasticity-fix — docs/benchmarks/e1-v3-locomo-results-post-fix.md.

Category	MRR	R@10
Adversarial	0.881	96.0%
Open-domain	0.875	96.9%
Multi-hop	0.779	90.3%
Single-hop	0.741	94.0%
Temporal	0.577	78.3%

No LLM at query time. Five signals fused — vector similarity, full-text search, trigram matching, thermodynamic heat, recency — then reranked by a cross-encoder. On PostgreSQL the fusion runs server-side in PL/pgSQL; on SQLite the same five signals are fused in-process.

BEAM — 10 million tokens of conversation

BEAM (Tavakoli et al., ICLR 2026) is the hardest long-term memory benchmark published: 10 conversations, each spanning 10 million tokens, probed across 10 memory abilities — including three no prior benchmark tests: contradiction resolution, event ordering, and instruction following. (Question counts vary by split: 196 on 10M, 395 on the current 100K.)

Every system in the paper collapses at this scale; the best reported (LIGHT on Llama-4-Maverick) scores 0.266 end-to-end. The collapse is measurable — and structured assembly resists it. Same code, same day, clean database, 35 conversations per split:

Split	Flat WRRF	With Context Assembler	Δ
500K (699 Qs)	0.500	0.570	+0.070
1M (695 Qs)	0.466	0.535	+0.069

Measured 2026-06-11 — benchmarks/results/beam_crossover/RESULTS.md. Flat retrieval degrades as the corpus doubles (0.500 → 0.466); the assembler holds a durable +0.07. At small scale it is net-flat (April 100K: 0.591 flat vs 0.602 assembled, 200-Q split since re-based to 395) — the value is scale-dependent, not universal.

At 10M tokens the gap widens — and the assembler needs no labels:

Configuration	MRR	vs. flat WRRF (0.353)
Flat WRRF baseline	0.353	—
Assembler, oracle stage labels (BEAM plan_id)	0.429	+21.5%
Assembler, temporal stage detection (timestamps only)	0.471	+33.4%

2026-04 family, same code revision, 196 Qs / 10 conversations — benchmarks/beam/variance/assembler_10m_stagefixed.txt and assembler_10m_temporal.txt. Reproduced 2026-06-11 on current code (fresh DBs, same 196 Qs): oracle 0.496, temporal 0.523 — the temporal advantage persists across code revisions (benchmarks/results/beam10m_paired/RESULTS.md).

The finding that surprised us: label-free temporal day-level partitioning outperforms BEAM's ground-truth topic labels (0.471 vs 0.429). Temporal proximity is a stronger stage signal than topic boundaries for conversational memory, so the Stage-Aware Context Assembly architecture deploys without any oracle metadata. It was originally designed in September 2025 for 9-page PRDs on Apple Intelligence's 4,096-token window (ai-prd-builder, commit 462de01) — one month before the BEAM paper existed — because the problem is the same at both scales: you can't fit everything in context, so you have to be smart about what goes in.

Honest caveat: BEAM defines no retrieval MRR metric — the paper uses LLM-as-judge nugget scoring. Our "MRR" is a retrieval proxy (rank of the first substring-matching memory); LIGHT's scores are end-to-end QA. The two are not commensurable, so we make no head-to-head BEAM claim and use BEAM only for within-system, same-harness comparisons.

pip install -e ".[postgresql,benchmarks,dev]"

python benchmarks/beam/run_benchmark.py --split 100K          # ~10 min
CORTEX_USE_ASSEMBLER=1 python benchmarks/beam/run_benchmark.py --split 10M
python benchmarks/locomo/run_benchmark.py                     # ~40 min
python benchmarks/longmemeval/run_benchmark.py --variant s    # ~45 min

All scores on a fresh database (DROP + CREATE per run), TRUNCATE between conversations, FlashRank preflight verified. Full methodology: docs/papers/research-post-context-assembly.md.

---

Context that survives compaction

Claude has a 200k/1M token context window. During long sessions, when it fills, it compacts: summarizes older messages, strips tool outputs, paraphrases instructions. Important nuance evaporates; decisions you anchored early dissolve into vague summaries.

Hippocampal Replay fixes this — named after the phenomenon where your brain replays important experiences during sleep to consolidate them. It treats compaction as "sleep" and replays what matters when Claude "wakes up." Before compaction hits, a hook drains your active context — what you were working on, which files were open, what decisions you'd made, what errors were unresolved — and stores it as a checkpoint. After compaction, a second hook reconstructs context intelligently: the latest checkpoint, anything you'd anchored as critical, the hottest project memories, and predictions about what you'll need next.

You can be explicit about what matters:

cortex:anchor({ content: "We're using event-sourcing. All state changes go through the event bus.", reason: "Architecture constraint" })

Anchored memories get maximum protection — they always survive compaction, no matter what.

The compaction checkpoint, session-start injection, and the autonomous wiki cycle are lifecycle hooks registered by the Claude Code plugin install. The single-click .mcpb bundle is a Directory connector — it delivers the 52 memory tools but no hooks (the MCPB format carries none). For the automatic session-lifecycle memory (session-start injection, auto-capture, compaction checkpointing, the autonomous wiki cycle), install the Claude Code plugin (see More options); the plugin also auto-registers the 3 upstream-integration tools when automatised-pipeline / prd-spec-generator are present (55 total).

---

The autonomous wiki

Cortex's wiki is a self-curating per-project knowledge base, not a memory dump. Every project the registry knows is driven toward 42 canonical documentation scopes (product overview, architecture, services, API, data flow, operations, decisions, onboarding, security, testing, configuration … ), and every source file toward 13 canonical sections (Purpose · Public API · Dependencies · Callers · How it works · Invariants · What can go wrong · Tests · Sequence diagram · Flow diagram · Parameters · Request example · Response example).

What makes it autonomous — no cron, no daemon, no manual invocation:

• A SessionStart hook spawns a background consolidate cycle every 6 hours (TTL stamp at ~/.claude/methodology/.last_consolidate). The agent runs because you opened Claude Code, and stops when nothing is left to author.
• A curation-gap detector + headless authoring worker. Each file-doc page declares its missing sections in frontmatter (curation_gaps:); the worker drains them by invoking claude -p (your existing credentials, no API key), which calls the codebase-intelligence MCP tools — codebase_context, codebase_impact, codebase_query — to ground each section in the real call graph before writing.
• Missing-anchor authoring. When a project has no architecture / services / api / data-flow / operations / ADR / PRD page, the worker authors it from the source tree (structure + README + manifest + CLAUDE.md), same grounding.
• Drift detection. Pages whose cited source moved, whose mtime is stale (>60 days), or whose body is off-template are flagged and re-authored in place. Deletion is never the policy; visibility is — a yellow banner shows ⚠ Page N% curated — M sections still missing and exactly what belongs in each.
• ADRs as task-records. Every completed task (≥1 commit at session end) auto-drafts an ADR with five mandatory sections (Entry / Mandatory / How / Result / Serves) from commit subjects + the session's memories; the worker refines it next cycle.
• Per-project dashboards at wiki/_dashboards/<project>.md show slot-fill rate, file coverage %, open gaps, and the queue for the next cycle.
• Continuous mechanical grooming. A recurring citation-reconciliation pass and a lesson-promotion backlog count run inside every consolidation cycle, not just when a session author happens to be active — closing a gap that let a 76-day, multi-thousand-item wiki-citation and lesson-promotion backlog go entirely unnoticed. get_grooming_health (a READ_ONLY MCP tool) reports exact backlog + last-run age per grooming type on demand; session_start prints one line when any type exceeds a sourced staleness threshold (3× the measured p90 gap between consolidation cycles). A scripts/groomer.py scheduled entry point (dry-run by default, opt-in --apply) is shipped for anyone who wants the maintenance pass to also run on its own cron/launchd schedule, independent of live sessions.

This isn't documentation you write — it's documentation Cortex authors, grooms, and verifies for you, every 6 hours, until every project reaches full scope coverage and every source file has all 13 sections filled.

Write papers in Cortex

Every page is editable in place in a full scientific writing environment — the same markdown that feeds the memory pipeline, with a rendering layer on top that never steals your content into a proprietary format. Your .md files stay grep-able, diffable, and git-versioned.

• CodeMirror 6 split-pane editor — syntax-highlighted markdown on the left, fully-rendered article on the right, atomic round-trip to the .md file on disk.
• Structured frontmatter — kind / domain / scope / status / authored_by / provenance / created / updated / last_reviewed. Real metadata: the coverage audit, dashboards, and wiki view all read it.
• [[wiki/path]] cross-references rendered as clickable links (bare slugs route to filtered search), with a backlinks footer.
• Mermaid diagrams with a 🔍 lens — viewport-sized viewer with wheel-zoom, drag-pan, and keyboard shortcuts.
• LaTeX math via KaTeX, BibTeX citations ([@friston2010] → (Friston 2010) with an auto APA bibliography), and figure / equation / table auto-numbering with cross-refs.
• Pandoc export — one click to PDF (via LaTeX), TEX, DOCX, or HTML. Journal-submittable from the same source.

The wiki's editor, galaxy graph, and views render through the standalone cortex-viz MCP, which reads this same store read-only.

---

Agent Integration

Cortex works with teams of specialized agents — and it uses one itself: the headless wiki worker is a Claude agent that drains the curation queue every six hours (see the autonomous wiki). Memory is shared across a team via Wegner's transactive-memory model (1987): teams store more than individuals because each member specializes.

• Specialization — each agent writes to its own agent_topic. Engineer's debugging notes don't clutter tester's recall; the wiki worker writes to agent_topic=wiki-curation so its drafts stay out of interactive recall.
• Coordination — decisions auto-protect and propagate. When engineer decides "use Redis over Memcached," every agent sees it at next session start. The ADRs the worker drafts at session end are the cross-agent shared memory.
• Directory — entity-based queries span all topics. "What do we know about the reranker?" returns results from engineer, tester, researcher, and the worker's drafts alike.

Works with any custom agents. See zetetic-team-subagents for a ready-made team of specialists, each with scoped memory.

---

The rest of the stack

Cortex fixes what an agent forgets. Three sibling MCP servers fix what it can't see, what it can't verify, and how it reasons. Each installs on its own — none requires the others — and Cortex registers extra tools automatically when it finds them.

	The problem it solves	Install it when	Related work
automatised-pipeline	Your agent answers structural questions ("who calls this?", "what breaks if I change it?") by re-reading files, burning context and missing cross-file callers. Indexes the repo into a property graph: call/import resolution, Leiden communities, hybrid BM25+TF-IDF search, impact analysis.	Refactoring, root-cause work, or any repo big enough that grep stops being an answer.	codebase-memory-mcp covers far more languages (158). automatised-pipeline is narrower but qualifies every impact answer as exact or lower-bound instead of presenting a possibly-incomplete list as complete.
prd-spec-generator	Specs pass review, then the implementation quietly contradicts them. Turns a feature description into a 9-file PRD and verifies it — deterministic Hard Output Rules plus multi-judge consensus calibrated against external oracles (schema / math / code).	You already write specs and want a gate that fails, not a template that hopes.	spec-kit, BMAD and Kiro generate specs. This verifies them — it runs as a CI gate over their output rather than replacing them.
zetetic-team-subagents	Subagents that assert confidently instead of citing. 11 problem-shaped skills over 97 sourced reasoning patterns (Curie to Toulmin), plus a pre-commit gate that blocks unsourced constants.	You want "I don't know" to be an available answer, and every constant in your code to trace to a paper or a benchmark.	Collections like wshobson/agents organise agents by role. These are organised by problem shape, and each carries its epistemic method and sources.
cortex-viz	Memory you can't inspect is memory you can't trust. Read-only galaxy graph, execution trace, and wiki browser over this same store.

---

Architecture

Clean Architecture with strict dependency rules — inner layers never import outer layers.

Layer	What lives here	Modules
shared/	Pure utilities (text, hash, similarity, types)	18
core/	Neuroscience + retrieval + wiki-curation logic	177
core/context_assembly/	Structured context assembler + stage detector	10
infrastructure/	SQLite + PostgreSQL stores, embeddings, file I/O, MCP client	59
handlers/	MCP tools + consolidation cycles (50 MCP-exposed; 53 with upstream integrations)	105
hooks/	Lifecycle automation (incl. autonomous consolidate spawn)	9 registered
server/	MCP tool registration + composition roots	—
observability/	Prometheus text-format metrics	2

Storage: SQLite by default (a single local file, zero setup) or PostgreSQL 15+ with pgvector (HNSW) and pg_trgm. Both back the same 52 tools and the same WRRF fusion of five signals — vector search, FTS, trigram, heat, recency. On PostgreSQL it runs server-side in PL/pgSQL stored procedures; on SQLite the equivalent fusion runs in-process (vector search included, as the live memory_stats has_vector_search flag confirms).

Concurrency (PostgreSQL): psycopg_pool.ConnectionPool with two latency classes — interactive_pool (min=2, max=8) for recall/remember/anchor, batch_pool (min=1, max=2) for consolidate/ingest. Tool handlers run on worker threads via asyncio.to_thread; per-tool admission semaphores bound fan-out. Heat is computed at read time by effective_heat(), so homeostatic maintenance writes one scalar per domain per run instead of N rows.

Configuration: select the backend with CORTEX_MEMORY_STORE_BACKEND (sqlite / postgresql / auto); set CORTEX_MEMORY_DATABASE_URL for the PostgreSQL path. All other parameters use the CORTEX_MEMORY_ prefix — see mcp_server/infrastructure/memory_config.py. Wiki cycle TTL is CORTEX_CONSOLIDATE_TTL_HOURS (default 6h).

---

Verification

Every benchmark headline above is backed by a per-mechanism ablation campaign — full n, single-seed, with code SHAs, dirty flags, manifests, and per-row JSON preserved:

• LongMemEval-S, 17 rows, n=500 — docs/benchmarks/e1-v3-results.md. Per-mechanism deltas at the calibrated equilibrium + category-specialization analysis.
• LoCoMo, 14 rows, n=1986 — docs/benchmarks/e1-v3-locomo-results.md (pre-fix) and docs/benchmarks/e1-v3-locomo-results-post-fix.md (post plasticity result-shape fix). Two-baseline design (NO_CONSOLIDATION / WITH_CONSOLIDATION).

The full per-mechanism evidence lives in the thermodynamic paper (§6.3); the BEAM decay dose-response (§6.4) documents a re-scoped negative result after a dirty-store confound was caught and traced. Thermodynamic Memory vs. Flat-Importance Stores (PDF, 34 pages) · Stage-Aware Context Assembly (PDF, 39 pages).

---

Security

Runs 100% locally — MCP over stdio, the storage backend (SQLite file or PostgreSQL on localhost) never leaves your machine (the optional cortex-viz companion binds its server to 127.0.0.1). No data leaves your machine. SafeSkill scan: 94/100 (code 97, content 88 — docs/safeskill-report.json).

Privacy Policy

Cortex is local-first: your memories, conversations, and profiles stay on your machine — stored in a local SQLite database (~/.claude/methodology/memory.db) by default, or in a PostgreSQL database you control. Cortex sends no memories, content, or telemetry to the author, Anthropic, or any third party. The only outbound network activity is a one-time download of open-source embedding/reranking models from Hugging Face (model files only), plus any integrations you explicitly configure. Full policy: PRIVACY.md.

Support

• Issues & bug reports: GitHub Issues
• Security disclosures: see SECURITY.md
• Contact: admin@ai-architect.tools

Development

pytest                    # full suite (see assets/badge-tests.svg for the current count)
ruff check .              # Lintruff format --check .     # Format
python scripts/check_doc_claims.py   # advertised counts must match the repo

Contributing: CONTRIBUTING.md · Who decides and what happens if the maintainer stops: GOVERNANCE.md · Where the project is going: docs/ROADMAP.md · The security argument and its limits: docs/ASSURANCE-CASE.md.

License

MIT — see LICENSE.

This software is the independent work of Clément Deust. It was developed outside any employment relationship and is not affiliated with, endorsed by, or owned by any past or present employer. It is part of the ai-architect ecosystem (zetetic-team-subagents, automatised-pipeline, prd-spec-generator).

The neuroscience and information-retrieval algorithms encoded in this software are derived from published academic work cited in docs/papers/bibliography.md and inline in the source via # source: annotations (Friston on predictive coding, Anderson & Lebiere on rate-distortion forgetting, Nader et al. on retrieval-induced lability, McClelland et al. on consolidation, and others). The MIT license covers this implementation; it does not assert ownership over the underlying mechanisms, which remain attributable to their original authors and publications.

Citation

The paper PDFs on main are the canonical artefacts (arXiv IDs forthcoming, endorsement in progress):

@software{cortex2026,
  title={Cortex: Persistent Memory for Claude Code},
  author={Deust, Clement},
  year={2026},
  url={https://github.com/cdeust/Cortex}
}

@unpublished{deust2026thermodynamic,
  title={Thermodynamic Memory vs. Flat-Importance Stores:
         Why Long-Term Retrieval Collapses Without Decay},
  author={Deust, Clement},
  year={2026},
  note={arXiv ID forthcoming, endorsement in progress},
  url={https://github.com/cdeust/Cortex/blob/main/docs/arxiv-thermodynamic/main.pdf}
}

@unpublished{deust2026context,
  title={Stage-Aware Context Assembly for Long-Context Memory Retrieval},
  author={Deust, Clement},
  year={2026},
  note={arXiv ID forthcoming, endorsement in progress},
  url={https://github.com/cdeust/Cortex/blob/main/docs/arxiv-context-assembly/main.pdf}
}