Question 1

Does every save and recall call go out to a language model?

Accepted Answer

Not on the default path, and that is the deliberate design point: the default profile makes no LLM call on write or search, which is why retrieval can be measured offline and a rerun gives the same result. LLM providers exist as a separate configuration layer, but they are opt-in rather than part of the profile the benchmarks describe.

Question 2

Is retrieval just vector similarity search?

Accepted Answer

No. Four retrievers run and are fused: full-text BM25, dense embeddings computed locally, fuzzy matching and a knowledge graph, combined with reciprocal rank fusion. An optional cross-encoder can rerank the fused list, and it is an install extra because it pulls in PyTorch, which tells you how much heavier the default path is.

Question 3

Will this stop my agent re-explaining the same conventions every session?

Accepted Answer

That is the stated purpose - decisions, solutions, facts, errors and session summaries are stored locally and served over MCP, so an agent can call memory_save and later memory_recall. But the evidence is retrieval-level: the documented numbers are recall-at-5 on public benchmarks. Nothing in the README measures whether an agent reliably acts on what it recalls.

Question 4

Any install traps I should know before wiring this into a client?

Accepted Answer

Three. Python 3.11 or newer is required, and CI only covers 3.11 to 3.13 on Linux, Windows and macOS. Memory lands in ~/.tam unless TAM_MEMORY_DIR points elsewhere, which matters for containerised setups. And the README says to use one of the two Claude Code plugins, not both.

Question 5

How do I register it with my coding agent?

Accepted Answer

Run tam setup at a terminal; the wizard asks just me or company server and detects Claude Code, Claude Desktop, Codex, Cursor, Windsurf, Gemini CLI, Cline and OpenCode, registering the server with the ones you pick. You can also add it by hand, for example claude mcp add memory -- total-agent-memory, then restart the client.

Question 6

Where does my memory live, and can I move it?

Accepted Answer

In ~/.tam by default, or wherever TAM_MEMORY_DIR points. Each store is one directory built around a SQLite database, and the Docker channel mounts ~/.tam at /data. The npm connector, Docker image and Claude Code and Codex plugins all reach the same store.

Question 7

Can several people share one memory store?

Accepted Answer

Yes, as a team server. Personal, department and company areas are separate stores behind one gateway, with roles, authorship and an audit trail. The PostgreSQL backend of the team server is a separate install extra, and the default local profile remains a single-directory SQLite store per person.

Total Agent Memory

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community