recall-echo 4.7.0

Persistent memory system with knowledge graph — for any LLM tool
docs.rs failed to build recall-echo-4.7.0
Please check the build logs for more information.
See Builds for ideas on how to fix a failed build, or Metadata for how to configure docs.rs builds.
If you believe this is docs.rs' fault, open an issue.
Visit the last successful build: recall-echo-4.1.0

recall-echo captures your agent's sessions as they end, turns them into a knowledge graph while the machine is idle, and gives every claim a confidence that rises with corroboration and falls with contradiction. Your agent asks it questions over MCP. The database and the embeddings run locally; only extraction calls a model, and that can be the CLI subscription you already pay for.

Install

curl -fsSL https://raw.githubusercontent.com/dnacenta/recall-echo/main/install.sh | sh
recall-echo init

init finds your agent CLIs (Claude Code, Codex, Grok, Gemini), installs the hooks, registers the MCP server with each and downloads the embedding model. With cargo: cargo install recall-echo --locked. Prebuilt binaries are on the releases page.

Using it

Day to day you do nothing. To look, or to fix something:

recall-echo what-do-you-know --about D          # what it believes, and how sure it is
recall-echo graph query "auth flow"             # search by meaning
recall-echo graph correct D USES Vim --wrong    # that claim is mistaken
recall-echo status                              # health

Inside the agent, it asks for itself:

You: why did we drop the websocket approach? Agent: (calls recall_query) You moved off it in June: the reconnect logic kept dropping messages under load, and you settled on polling every 30 seconds.

On a LongMemEval subset it went from 0 of 9 answers right to 7 of 9. A small sample, stated as one: benchmarks.

Docs

License

MPL-2.0. Versions up to v3.13.0 remain available under AGPL-3.0. The graph store, SurrealDB 3.x, is BSL 1.1: embedding it is allowed, and it becomes Apache-2.0 in 2030.