Skip to content

senna-lang/Codeatrium

Repository files navigation

codeatrium — two commands, recall everything: a minimal memory layer for AI coding agents

CI PyPI License: MIT

English · 日本語

loci search recalling a past design decision with its symbol, file:line, and git branch

An AI coding agent recalls everything it has done through just two commands: loci search and loci context. That's the whole interface. The agent reaches for the right call without hesitation, and restores past decisions, conversations, and exact code locations in under 0.2 seconds.

The CLI command loci is designed to be called by the agent itself — running loci search "..." --json from within a prompt. (The name comes from the Method of Loci — the memory-palace technique. Under the hood, conversations are distilled into "palace objects"; see How It Works. The architecture extends the conversational memory model from arXiv:2603.13017 for coding agents.)

Note: Currently Claude Code only — indexing reads Claude Code's session logs (.jsonl). Distillation defaults to claude --print (Haiku) but can run on a local OpenAI-compatible LLM instead (see Configuration).

Minimal Interface

The whole recall interface is two commands:

  • loci search "query" — semantic search over past conversations
  • loci context — reverse lookup, by code symbol (--symbol "name") or git branch (--branch "name")
    • tree-sitter symbol resolution (Python / TypeScript / Go) lets agents understand implementation intent before editing
    • --branch "name" recalls what was done and discussed on a specific git branch (also available as loci search "query" --branch "name")

That's deliberate. The user here is the agent, and an agent handed a 50-tool palette hesitates, mis-picks, and burns tokens just deciding which to call. With a surface this small — and no MCP tool schemas sitting resident in the context window — the agent reaches for the right call the first time, every time. (When the full transcript is needed, loci show "<ref>" expands any result to its verbatim source.)

Touching a symbol means recalling what was decided about it — loci context reverse-looks-up the exact code location, signature, and the conversation behind it:

loci context reverse-looking-up a symbol to the conversation that shaped it

How It Works

session logs are indexed into exchanges, distilled into palace objects with symbols, then recalled via BM25 + HNSW fused by RRF

  1. Index — Splits agent session logs into exchanges (user utterance + agent response pairs) and indexes them with FTS5 for keyword search
  2. Distill — An LLM (claude --print, default claude-haiku-4-5) summarizes each exchange into a palace object: exchange_core (what was done), specific_context (concrete details), room_assignments (topic tags). tree-sitter resolves touched files to symbol level (function/class/method + file + line + signature)
  3. Search — Cross-layer search fusing BM25 on verbatim text with HNSW on distilled embeddings via RRF

Raw conversations are not embedded — only the condensed distilled text is embedded with multilingual-e5-small (384-dim), balancing semantic search quality with embedding cost. The embedding model runs as a Unix socket server, keeping search latency under 0.2 seconds after the first load.

Installation

pipx install codeatrium

Requires Python 3.11+.

Quick Start

# Initialize in project root (also registers Claude Code hooks)
loci init

loci init now handles everything in one step — database setup, existing session detection, and Claude Code hook registration. Pass --no-hooks to skip hook registration. If init fails partway through, .codeatrium/ is cleaned up automatically so re-running is safe.

When running loci init, if past session logs are detected, you'll be prompted with:

Important

When adopting this tool mid-project, a large number of exchanges may already exist. Distilling all of them will consume significant claude --print (Haiku) tokens. We recommend starting with Skip all or Distill last 50.

  1. Min chars threshold — Minimum character filter applied at index time (default: 50). Shorter exchanges are skipped entirely, which also shrinks the pool of distillation candidates. Higher values exclude short conversations and reduce token usage; lower values include nearly everything. (Distillation applies a separate min_chars of 100 — see Configuration.)
  2. Handling existing exchanges — Choose how much past history to distill:
    • Skip all (no past session distillation)
    • Distill last 50 (recent history only)
    • Distill all (everything — high token cost)
    • Custom (specify a number)
  3. Run distillation now? — Accepts 1/2/y/n/yes/no. Choose No to defer to the next session start.

Invalid input on any prompt re-prompts instead of silently falling back to a default.

Agent Instructions

Agent instructions are injected automatically — no manual setup required:

  • loci init — Inserts a marker section (<!-- BEGIN CODEATRIUM -->...<!-- END CODEATRIUM -->) into CLAUDE.md
  • loci prime — Dynamically injects command usage into the context window at every session start via SessionStart Hook

CLI Commands

Command Description
loci init Initialize .codeatrium/ and register Claude Code hooks (--no-hooks to skip)
loci index Index new session logs
loci distill [--limit N] Distill undistilled exchanges via LLM
loci search "query" --json Semantic search (agent-facing); add --branch NAME to filter by git branch
loci context --symbol "name" --json Code symbol → past conversations (lightweight; add --full for verbatim text)
loci context --branch "name" --json Git branch → past conversations (includes undistilled exchanges)
loci show "<ref>" --json Retrieve verbatim conversation
loci status Show index state
loci prime Inject command usage into the session context (run automatically by the SessionStart hook)
loci server start/stop/status Embedding server management
loci hook install Re-register hooks (normally already done by loci init)
loci hook uninstall Remove codeatrium hooks from settings.json

Automation (Claude Code Hooks)

After loci init (or loci hook install), everything runs automatically:

Hook Trigger Command
Stop (async) After every turn loci index
SessionStart startup / /clear / /resume / compact loci prime
SessionStart startup / /clear / /resume / compact loci server start
SessionStart startup / /clear / /resume / compact loci distill
  • loci index — Runs asynchronously after every turn. Indexes only new exchanges, so it's fast even mid-session
  • loci distill — Distills undistilled exchanges at session start. Defaults to claude --print (Haiku, through the user's Claude Code); can use a local OpenAI-compatible LLM instead (see Configuration)
  • loci server start — Keeps the embedding model (~500MB) resident in memory for sub-0.2s search latency

Search Output

[
  {
    "exchange_core": "Added connection pool with pool_size=5",
    "specific_context": "pool_size=5, max_overflow=10",
    "rooms": [
      { "room_type": "concept", "room_key": "db-pool", "room_label": "DB connection pooling" }
    ],
    "symbols": [
      { "name": "create_pool", "file": "src/db.py", "line": 42, "signature": "def create_pool(...)" }
    ],
    "verbatim_ref": "~/.claude/projects/.../session.jsonl:ply=42",
    "git_branch": "feature/db-pool"
  }
]

Configuration

.codeatrium/config.toml (generated by loci init):

[distill]
provider = "claude"                    # Distillation backend: "claude" | "openai" (default "claude")
model = "claude-haiku-4-5"             # Model for distillation (default)
batch_limit = 20                       # Max distillations per hook run
min_chars = 100                        # Skip distillation for exchanges shorter than this

[index]
min_chars = 50                         # Skip indexing exchanges shorter than this

There are two min_chars settings: [index] min_chars controls what gets indexed at all, while [distill] min_chars further skips distillation (the LLM cost) for short exchanges that were already indexed.

Distilling with a local LLM

Distillation is a small per-exchange structured-extraction task, so a local model is usually good enough. Any OpenAI-compatible endpoint (Ollama, LM Studio, llama.cpp-server, vLLM) works by setting provider = "openai" and base_url — no new dependencies, no API key (the Authorization header is never sent, so this is local-only):

[distill]
provider = "openai"
model = "qwen2.5:7b"
base_url = "http://localhost:11434/v1"   # Ollama
# base_url = "http://localhost:1234/v1"  # LM Studio

base_url is required when provider = "openai"; if it is missing or empty the provider falls back to claude with a warning. With provider = "claude" (the default), base_url is ignored and distillation runs through claude --print as before.

Acknowledgments

The palace object model, room-based topic grouping, and BM25+HNSW fusion search are based on:

Structured Distillation for Personalized Agent Memory (arXiv:2603.13017)

License

MIT

About

Minimal local memory for AI coding agents — two commands, recall everything.

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors