memory

Persistent, local memory for AI coding agents: Claude Code, Codex CLI, Cursor and any MCP client.

Website and documentation: totalmemory.dev · Quick start · Claude Code guide · Benchmarks · Compare with other tools

total-agent-memory (TAM) is an open-source memory server for AI coding agents. Coding agents start every session without memory of earlier ones, so decisions, fixes and project conventions have to be explained again. TAM stores decisions, solutions, facts, errors and session summaries on your machine and returns them through the Model Context Protocol (MCP), so any MCP client can use it without code changes.

Each store is one directory built around a SQLite database. Recall combines full-text BM25, dense embeddings computed locally, fuzzy matching and a knowledge graph, and fuses the ranked lists with reciprocal rank fusion. A local cross-encoder then reranks the result: by default it starts as soon as its model has loaded in the background, and MEMORY_CROSS_RERANK=off turns it off. The default profile makes no LLM call on write or search, so retrieval can be measured offline. With the cross-encoder off, reruns on fresh stores returned the same ranking for every LoCoMo question. With the default setting the first queries of a process can run before the model has loaded, so reruns are not guaranteed to match. Facts can carry validity intervals (kg_add_fact, kg_at), and a newer value of a single-valued fact can retire the older one.

TAM is for developers who use coding agents daily and for researchers who need a memory baseline they can run locally, inspect and change. The same package can also run as a team server: personal, department and company areas are separate stores behind one gateway, with roles, authorship and an audit trail (team server).

Installation

TAM needs Python 3.11 or newer. CI tests Python 3.11, 3.12 and 3.13 on Ubuntu, Windows and macOS.

From PyPI (use a virtual environment, or pipx / uvx):

pip install total-agent-memory          # or: pipx install total-agent-memory
uvx total-agent-memory                  # run once without installing

This installs the total-agent-memory (alias tam) MCP server and the tam-team, tam-remote and lookup-memory commands. The default cross-encoder runs on ONNX through fastembed and needs no extra. A separate sentence-transformers reranker pulls in PyTorch and is an extra: pip install "total-agent-memory[rerank]". The PostgreSQL backend of the team server is the [postgres] extra.

Register it with your clients. Run tam setup at a terminal. The wizard asks "Just me" or "Company server", detects Claude Code, Claude Desktop, Codex, Cursor, Windsurf, Gemini CLI, Cline and OpenCode, and registers the server with the ones you pick (setup wizard).

First start. With no command, tam starts the MCP server on stdio, which is how MCP clients run it. tam --help lists the commands and tam --version prints the version (from 14.11; earlier versions start the server for any argument, so use tam setup --help there). The first start downloads about 330 MB of models: the text embedder (about 240 MB) and the cross-encoder (about 90 MB). The cross-encoder loads in the background, so a short script that exits after a few queries gets the fused order without it. fastembed keeps the models in the system temp directory, which macOS clears from time to time; set TAM_MODEL_CACHE=~/.tam/models to keep them.

Other channels:

Use one of the two Claude Code plugins, not both. Per-platform details, WSL2, the IDE matrix, uninstalling and troubleshooting are in the installation guide.

Quick start

1. Install the package and run tam setup, or add the server to your client by hand. For Claude Code:

   claude mcp add memory -- total-agent-memory
   

   For clients that use an mcpServers file:

   { "mcpServers": { "memory": { "command": "total-agent-memory" } } }
   

   If you installed into a virtual environment, use the full path to its total-agent-memory executable.

2. Restart the client. Memory is stored in ~/.tam/ unless TAM_MEMORY_DIR points elsewhere.

3. Ask the agent to remember something ("remember that we chose PostgreSQL for billing because of row-level security"). It calls memory_save. In a later session, ask "which database did we choose for billing?"; it calls memory_recall and gets the record back.

To try the server without an agent, this script starts it over stdio with the MCP Python SDK (installed as a dependency), saves one record and recalls it. The SDK passes only a few variables to the server by default, so the script hands over the full environment, including TAM_MEMORY_DIR:

import asyncio
import os

from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client


async def main() -> None:
    server = StdioServerParameters(command="total-agent-memory", env=dict(os.environ))
    async with stdio_client(server) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()
            await session.call_tool("memory_save", {
                "type": "decision",
                "content": "Chose PostgreSQL over MySQL for the billing service",
                "context": "WHY: row-level security per tenant",
                "project": "demo",
            })
            result = await session.call_tool("memory_recall", {
                "query": "which database for billing", "project": "demo", "limit": 3,
            })
            print(result.content[0].text)


asyncio.run(main())

TAM_MEMORY_DIR="$(mktemp -d)" python quickstart.py

The first run downloads the default embedding model. The output is JSON with the saved decision under results.decision. More examples: tools and interfaces.

Running the tests

The tests run from a source checkout. These commands mirror the CI workflows in .github/workflows:

git clone https://github.com/vbcherepanov/total-agent-memory.git
cd total-agent-memory
python3 -m venv .venv
.venv/bin/pip install -e . -r requirements-dev.txt

export FASTEMBED_CACHE_PATH="$PWD/.tam-models" TAM_MEMORY_DIR="$PWD/.tam-test-memory"
.venv/bin/python tests/smoke/prewarm_models.py   # downloads the text and code embedding models and the cross-encoder (~940 MB) once
.venv/bin/python -m pytest tests -q

The tests need no API keys and no LLM; the full suite takes 15 to 20 minutes on a laptop. Do not set MEMORY_LLM_ENABLED=false for the full suite: the configuration tests check LLM auto-detection. Tests that need services you do not have are skipped:

• PostgreSQL (team server backend): install the extra with .venv/bin/pip install -e ".[postgres]", then run .venv/bin/python -m pytest tests -m postgres --backend=postgres or --backend=both. Without TAM_TEST_PG_URL the fixtures start a pgvector container through Docker. See CONTRIBUTING.md and .github/workflows/postgres.yml.
• Browser tests (tests/browser, Playwright with Chromium, Firefox and WebKit) run in the image built from docker/Dockerfile.browser; see the browser job in .github/workflows/smoke.yml.
• Installed-package smoke test: python tests/smoke/installed_runtime.py checks an installed wheel over local and remote MCP.

Reproducing the benchmarks

The retrieval benchmarks (LoCoMo, LongMemEval, BEAM) need no API key and run with scripts in benchmarks/ once the public datasets are downloaded. The table below was measured on v13.0.0, before the cross-encoder became the default in 14.5, so it describes retrieval without the cross-encoder and not a default install of the current version:

Default settings score higher on LoCoMo and take much longer per query. In the JOSS review of v14.6.0 the editor measured R@5 of about 0.745 at about 495 ms median with the defaults, against 0.607 at about 18 ms with MEMORY_CROSS_RERANK=off. Our own runs of v14.6.0, three on a fresh store per setting, gave R@5 0.611 with the cross-encoder off and 0.746 with the default and with on, at about 0.95-1.0 s median per query on an Apple M2 Max; the median depends on the CPU. Each setting returned the same ranking for every question in all three runs. With the default, the first query of a process can arrive before the cross-encoder has loaded and keeps the fused order; in these runs that changed one question out of 1,986 (rows, reports and commands).

Dataset locations, commands, end-to-end (LLM-judged) accuracy, negative controls, latency and the 14.x studies are in