Doolta

Mnemos

Local, verifiable memory for AI agents: every answer cites its exact source, with no vector database and no heavy infrastructure.

Project Details

Real problemAn AI agent (Claude Code, for instance) forgets project context from one session to the next: it misses your ADRs, invents answers instead of reading existing documentation, and can’t reliably cite where a claim came from.
ConstraintsStay strictly local-first: no data leaves the machine, no vector database to operate, no Ollama, no Python or Node service to maintain. Produce verifiable answers rather than plausible ones. Let the agent write safely: read-only by default, writes confined to a path, content scanned for secrets.
Structural choicesA single, cgo-free, self-contained Go binary embedding an MCP server, an indexing pipeline, SQLite storage, and full-text search. A plain-Markdown knowledge format: any folder of Markdown files works as-is. Every result cites the exact file, section, and line range. Lexical search (SQLite FTS5/bm25) by default, optional semantic and hybrid search behind a build flag.
What it proves todayA published, continuously tested, versioned, MIT-licensed tool. Retrieval quality is measured, not just claimed: an evaluation command automatically derives query/source pairs from a knowledge base and reports retrieval metrics (default lexical search at 0.83, hybrid mode at 1.00 on the shipped example set).
Associated offerProof of capability in local-first, verifiable AI knowledge systems: bespoke design on request. Get in touch.

A memory the agent can cite, not invent

A well-equipped AI agent still forgets everything between sessions: it doesn’t know why an architecture was chosen, misses decisions already made, and fills the gaps with plausible but unsourced answers. Mnemos closes that gap by giving the agent a local, indexed, citable memory: every answer points to the exact file, section, and lines it came from.

Zero infrastructure, one dependency: your own files

No vector database to operate, no Python or Node service to maintain, no telemetry. A single, self-contained Go binary embeds the MCP server, indexing, storage, and search. Any folder of Markdown documentation becomes a usable knowledge base as-is.

Search quality, measured rather than promised

Instead of just claiming search works, the project ships its own evaluation: query/source pairs are automatically derived from an existing knowledge base, then used to measure actual retrieval accuracy. The same standard of verifiability applied to the agent’s answers is applied here to the search engine itself.

This project demonstrates a directly transferable capability: designing AI systems where every claim stays traceable to its source, without relying on heavy infrastructure or an external provider.