fridai 🛠️

English | 한국어

A lightweight MCP server that lets coding agents (Claude Code, etc.) recall your past code, commits, and AI conversations — and remember durable notes — 100% local. Search & recall only — no local LLM required.

• Search-only (read-only). The calling agent (an LLM) does the reasoning; fridai just returns evidence with sources.
• Local embeddings (fastembed, onnx). No Ollama or external API. Semantic search works out of the box (falls back to lexical if fastembed isn't installed).
• Multi-agent history. Auto-indexes Claude Code · OpenAI Codex CLI · Gemini CLI conversations.
• Private by design. Your memory never leaves the machine; secrets are auto-masked at index time.

Status: early development. Requires Python 3.10+ and git (for code/commit indexing).

How it works

1. Index (fridai index) builds a local database from three source types:
   • AI conversations — parses each agent's session logs into question→answer turns, and matches a question to its resulting git commit (time window ∩ touched files).
   • Code — chunks git-tracked files by function/class (line-window fallback), with line ranges.
   • Commits — indexes git commit history (subject + changed files).
2. Everything is stored in sqlite + FTS5 (lexical) plus float32 vectors at ~/.fridai/index.db.
3. Recall fuses lexical (BM25) and vector (cosine) results via RRF, then reranks so real work artifacts (code/commits/edited turns) outrank bare question turns, and dedupes repeated questions.

Install

pipx install fridai        # isolated (recommended)
# or: pip install fridai   # in a venv

Pulls numpy + fastembed (onnx) + mcp. Registers the fridai command.

Quickstart

fridai index --source all             # build the index (current repo + all agent conversations)
fridai stats                          # index overview
claude mcp add fridai -- fridai mcp   # register with Claude Code

After registering, the agent recalls via the recall tool. Re-run fridai index anytime to refresh — it's incremental (only changed files/sessions/new commits are reprocessed), so it's cheap. To keep it fresh automatically, run fridai index --watch (reindexes every 15s; --interval to change), or fridai install-hook to reindex on every git commit (no running process needed).

CLI reference

index flags:

MCP tools

The server exposes two tools — recall (read) and remember (write):

recall(query, k=5, repo="", source_type="")

• query — search text (natural language and/or code identifiers).
• k — max results (default 5).
• repo — empty = current working repo (server detects cwd); "all" = every repo; "<name>" = a specific repo.
• source_type — empty = all; one of agent_turn · code · commit · note.

It returns text with numbered, cited hits for the agent to read — for example:

fridai recall — 2 memory item(s) for "docker mount" (all repos, with sources):

### [1] myrepo 2026-07-01 [codex] session:how did I add the docker mount?
Q: how did I add the docker mount?
A: added a bind mount via volumes.

### [2] myrepo/docker-compose.yml:1-20
volumes: ...

remember(text, repo="")

The write counterpart to recall. The agent calls it mid-session to persist a durable note — a decision and its rationale, a non-obvious gotcha, or a recurring fix — that a future session would want back. repo defaults to the current working repo (so recall finds it there); pass "<name>" to pin it elsewhere. Notes are redacted and embedded like any other source.

Humans can write the same notes from the terminal with fridai note "..." (see the CLI reference).

Non-Claude sources are tagged (🤖 codex / 🤖 gemini in the CLI, [codex]/[gemini] in citations).

Registering with other MCP clients

fridai is a standard stdio MCP server — the launch command is fridai mcp. Any MCP-capable client can use it. For clients that read an mcpServers config:

{
  "mcpServers": {
    "fridai": { "command": "fridai", "args": ["mcp"] }
  }
}

fridai mcp --print-config prints ready-to-paste snippets for Claude Code, Codex CLI, and Gemini CLI (using the resolved absolute path, so GUI clients that don't inherit $PATH still work). Add --client claude|gemini|codex to print just one.

Indexed agent sources

A missing agent directory is silently skipped.

Security

Secrets (AWS keys, GitHub/Slack tokens, PEM private keys, JWTs, password=…) are auto-masked at index time (on by default). Nothing is sent off the machine.

fridai index --source code --no-redact   # disable masking
echo "*.env" >> .fridaiignore                 # exclude paths (repo root or ~/.fridai/)

The high-entropy heuristic is off by default (too many false positives on long identifiers); enable it with FRIDAI_REDACT_ENTROPY=1.

Environment variables

Embedder consistency: index and query with the same embedder. Mixing models makes vectors incomparable even at the same dimension — the store then falls back to lexical search. Reindex (--reindex) after switching models.

Non-English content: the default nomic-ai/nomic-embed-text-v1.5 is English-centric, so semantic recall on other languages (e.g. Korean) is weaker. For multilingual code/conversations, point FRIDAI_FASTEMBED_MODEL at a multilingual model and reindex — e.g. sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 (384-dim, small and fast, good all-round) or intfloat/multilingual-e5-large (1024-dim, higher quality, heavier). After changing it, run fridai index --reindex --source all.

Development

PYTHONPATH=src python3 -m unittest discover -s tests

CI (GitHub Actions) runs the suite on Python 3.10–3.14 plus a fastembed smoke check on every push/PR. Tests are hermetic — tests/__init__.py isolates all FRIDAI_* paths and disables the embedder, so no real ~/.fridai/~/.codex/etc. is touched and no model is downloaded.

License

MIT.