total-agent-memory

total-agent-memory

Persistent memory for your facts, decisions and working practices. Persistent, local memory for AI coding agents: Claude Code, Codex CLI, Cursor, any MCP client. Temporal knowledge graph · procedural memory · AST codebase ingest · cross-project analogy · 3D WebGL visualization.

Why this, not mem0 / Letta / Zep / Supermemory / Cognee? → docs/vs-competitors.md

---

Version 14.3.1 — updates are no longer dropped, recall stays fast at 1M records

Release date: 2026-09-21.

Two problems that got worse as a store grew. Dedup treated near-identical texts as repeats, so an update that changed one value ("Messi's citizenship is Argentina" → "... Armenia", "PostgreSQL 16" → "18") was dropped and the old value kept; on MemoryAgentBench FactConsolidation 14.3.0 lost 36 of 455 facts. And several queries behind every recall and save read the whole store. Measured on one machine with 200 synthetic tenants (report):

	14.3.0	14.3.1
Tenant-scoped memory_recall, 100k records, p50 / p95	781 / 1,901 ms	25 / 35 ms
Tenant-scoped memory_recall, 1M records, p50 / p95	6–64 s	105 / 147 ms
HTTP calls/s, 100k records, 16 clients	1.3	37 (one process), 117 (MCP_HTTP_WORKERS=4)

Change	How you use it
Updates are kept	Automatic. A record is a repeat only when it has the same words in the same order (case, punctuation and ё/е aside). A repeat replaces the stored record, so "A", then "B", then "A" leaves A as the latest.
Scoped search follows the project, not the store	Automatic. Migration 035 puts the project into the full-text index (rebuilt once, about 30 s per million records); graph seeds and available_solutions use indexes.
HTTP workers	MCP_HTTP_WORKERS=4 with MCP_TRANSPORT=http: four server processes on one port, stateless sessions, POSIX only. In Docker Compose: TAM_MCP_WORKERS=4.
Scale benchmark	benchmarks/scale_bench.py loads a synthetic multi-tenant corpus through the real save path and measures recall, save, HTTP throughput and concurrent writers.
MemoryAgentBench	FactConsolidation results: findings.

---

Version 14.3.0 — Jev as the contradiction checker

Release date: 2026-09-21.

memory_answer can now check retrieved facts for contradictions with TypeSafe's Jev, a System One model that answers typed questions with calibrated probabilities instead of generating text. Every (supporting, opposing) pair becomes one noul question, and all of them go in a single request. Measured on the same questions against the Claude Haiku 4.5 scorer (report): accuracy is unchanged within noise (LongMemEval knowledge-update 35/78 vs 36/78, control 15/50 vs 16/50). The median contradiction pass drops from 3.1 s to 1.9 s, and Jev billed $0.038 for all 78 questions.

Change	How you use it
Jev contradiction scorer	MEMORY_CONTRADICTION_SCORER=jev plus TYPESAFE_API_KEY. The default stays llm. The client retries 408, 429, 5xx and connection errors like the official SDKs, honours retry-after-ms / Retry-After up to 60 s, and reports jev_* counters and a jev_request_ms latency histogram.
API key hygiene	An empty key, or one containing a newline or other control character, is rejected before any request, and the error never contains the key. (The official Python and JS SDKs echo it; reported upstream as typesafe-sdk-python#9 and typesafe-sdk-js#14.)
Benchmark harness	benchmarks/knowledge_update_eval.py records per-question contradiction-pass time and Jev token usage, so the two scorers can be compared on cost as well as accuracy.

---

Version 14.2.0 — facts that change over time

Release date: 2026-09-21.

"Mary loves red", saved in May, and "Mary no longer likes red; she has fallen for green", saved in August: memory_answer now says green, previously red. Measured with Claude Haiku 4.5 on the same questions (report): LongMemEval knowledge-update rose from 12/78 to 35/78, a Russian + English update suite from 8/30 to 22/30, and five other LongMemEval categories from 11/50 to 16/50.

Change	How you use it
Recording dates reach the reader	memory_answer's reader and verifier see when each record was saved. The latest record about the same subject gives the current value unless its text describes the past, and a value stays current until a later record changes it. Answers come back in the language of the question.
Contradictions are resolved, not refused	A hard contradiction now passes both sides, with their dates, to the reader. MEMORY_CONTRADICTION_POLICY=abstain restores the 14.1.0 refusal. The contradiction scorer sees the question, so a conflict about someone else no longer blocks the answer.
Russian word forms	The lexical recall tier stems Cyrillic terms (Snowball, new dependency snowballstemmer): "Маша" in a question finds "Маше" in a record, and claim grounding accepts the inflected name.
One timestamp format	Records are stored as 2026-09-21T08:21:37.622445Z (UTC). Migration 034 rewrites older rows: +00:00 and fraction-less values are reformatted, and zone-less values, which older versions wrote in local time, are converted with that zone's DST rules. The instants do not change, so atomic facts are not rebuilt.

---

Version 14.1.0 — what is new

Release date: 2026-09-21.

Change	How you use it
Negative retrieval in memory_answer	Before reading, the grounded reader runs a second, contradiction-seeking search: a small model inverts the question, and each (supporting, opposing) pair — at most 5 × 5 — is scored in one batched call. Score ≥ 0.60 answers Not enough information without picking a side; 0.30–0.60 answers with a caveat; below 0.30 the answer is unchanged. The verdict is returned under negative. MEMORY_NEGATIVE_RETRIEVAL=false turns it off.
memory_answer on Anthropic and Ollama	Both providers now return schema-constrained output (forced tool call / JSON-schema format). Before this, Claude Haiku wrapped JSON in a markdown fence and memory_answer failed with Reader returned invalid grounded evidence.

---

Version 14.0.0 — what is new

Release date: 2026-09-15 · Status: release candidate; registry publication pending. Use the prepared wheel or this source checkout for v14. The general package-manager commands below follow their published channels and do not guarantee v14 before publication.

Change	How you use it
Personal, team and shared memory	Install one server; give Vasya and Petya separate tokens. Select a scope when saving; search all areas you can access. Each area has its own database, graph and index.
Authorship and revision history	The token identifies the author and client. See who changed a record, when and why; revision checks prevent overwriting another person's edit.
Remote MCP and team web interface	Connect an IDE through the lightweight Python bridge. In the browser, select a scope, search, browse, save, edit and inspect history. The interface displays the product name, version and release date.
Lower CPU pressure	Fast mode remains the default. Embedding and optional PyTorch models default to one compute thread. The team server retains three workspace workers to avoid repeated model loading.
Configurable internal LLMs	Use Ollama, an OpenAI-compatible endpoint or Anthropic for internal text tasks; explicitly select a vision model for images.
Safer retrieval and answers	Scoped context, model-aware vector search and privacy-safe write intents. The optional grounded reader checks supporting evidence and rejects contradictory claims; it is not the default answer path.
Installation and packaging fixes	Wheel, source archive and Docker checks cover Linux, Windows and macOS; the source archive now includes test fixtures and installation support files.

Validation: 2,162 tests passed in the checkout; 2,145 passed from the source archive. Native wheel checks passed on Ubuntu, Windows and macOS; 12 browser scenarios passed across Chromium, Firefox and WebKit. The expanded CI matrix still requires execution. See the final verification report for skips, exact platforms and artifact hashes.

Performance limits: BGE is optional. On the measured Linux ARM64 setup, its one-thread p95 was about 2,179 ms, above the 200 ms target. Limiting threads reduces parallel CPU load; it does not eliminate CPU work. No top-10 ranking or new default answer-quality improvement is claimed. CPU measurements.

Full v14 release notes · Changelog and previous releases

Getting started with v14

Choose local or server mode

Mode	Install and use
Local, one person	Follow native or Docker installation, then Quick start. Memory and models run on your machine; the local MCP catalogue has 74 tools.
Server, multiple people	Follow the setup below. Memory and models run on the server; clients need only Python and their token. The remote catalogue exposes eight core tools.

The server instructions work with the prepared v14 wheel on Linux, macOS and Windows. Run commands in a directory where you can create the data folder and token files.

Start a server without Docker

Install the candidate wheel in a dedicated environment. Linux/macOS:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install ./dist/total_agent_memory-14.3.1-py3-none-any.whl

Windows PowerShell, using the environment directly without changing the execution policy:

py -3 -m venv .venv
.\.venv\Scripts\python.exe -m pip install .\dist\total_agent_memory-14.3.1-py3-none-any.whl
$env:PATH = "$PWD\.venv\Scripts;$env:PATH"

Then, on any of these systems: