CodeMira

A developer memory system that learns from your coding sessions and surfaces relevant context when you need it.

CodeMira watches your OpenCode sessions, extracts patterns and decisions from idle conversations, stores them in a local knowledge base, and injects relevant memories into the LLM context on every call — so your AI assistant remembers what you've already figured out.

How It Works

Two processes cooperate:

1. Python daemon — runs in the background, monitors OpenCode sessions for idle time, compresses tool call transcripts and extracts memories via OpenAI-compatible LLM endpoints (Ollama, OpenRouter, vLLM, llama.cpp, OpenAI itself, etc.), stores them in SQLite + hnswlib (ANN) + FTS5, and serves retrieval requests over HTTP.

2. TypeScript plugin — hooks into OpenCode's experimental.chat.messages.transform, calls an OpenAI-compatible LLM (the "subcortical", Ollama by default) to analyze the current conversation intent, queries the daemon for relevant memories, and injects a <developer_context> HUD block into the message stream.

Retrieval uses hybrid search: BM25 (full-text), ANN (cosine similarity via hnswlib), and Reciprocal Rank Fusion. Entity-based hub discovery pulls in memories linked to the same technologies. A dedup layer (fuzzy text + vector similarity) prevents storing near-duplicates.

Storage Model

Memories are scoped per-project. There is no global store.

Each project gets its own memory store at <project-worktree>/.codememory/memories.db, alongside memories.index (hnswlib ANN cache rebuildable from the DB). Commit .codememory/ to share institutional knowledge with collaborators, or add it to .gitignore to keep memories local.

OpenCode itself uses a single global SQLite database (typically ~/.local/share/opencode/opencode.db) for all its session history, discriminated by project_id. The daemon reads that DB read-only, joins session.project_id against project.worktree to resolve each session's project root, and routes extracted memories into that project's .codememory/ directory. There is no shared memory namespace across projects — a preference learned in project A does not bleed into project B.

Install

Daemon

cd daemon
pip install -e .

Requires:

• Python 3.12+
• An OpenAI-compatible LLM endpoint for the subcortical / consolidation / arc roles. Default: Ollama at http://localhost:11434/v1 (Ollama 0.1+ exposes the OpenAI-compat path natively). Any provider that speaks POST /chat/completions works — vLLM, llama.cpp's HTTP server, LM Studio, etc.
• An OpenAI-compatible endpoint with API key for the extraction role. Default: OpenRouter (https://openrouter.ai/api/v1). Substitute OpenAI, Together, Anthropic-via-proxy, or your own self-hosted endpoint as needed.
• SQLite, hnswlib, rapidfuzz (installed via pip)

Set environment variables:

export CODEMIRA_EXTRACTION_API_KEY=sk-or-v1-...

Run:

python -m codemira.daemon

Or install as a macOS launchd service:

python -m codemira.launchd install

Plugin

Add the plugin to your global OpenCode config at ~/.config/opencode/opencode.json:

{
  "plugin": ["file:///path/to/CodeMira/plugin/src/index.ts"]
}

This makes it available in every OpenCode session regardless of working directory. OpenCode auto-discovers plugins from this config on startup.

Requires Bun or a bundler that handles TypeScript ESM. The plugin calls an OpenAI-compatible LLM for the subcortical model — by default Ollama at http://localhost:11434/v1. Override via subcorticalBaseUrl / subcorticalApiKey in plugin options to point at any other provider.

Prompt templates are loaded from prompts/ relative to the plugin source (plugin/src/../../prompts). Missing files raise a fatal error at plugin load — the plugin does not silently no-op.

Configuration

Variable	Default	Description
CODEMIRA_HTTP_PORT	9473	Daemon HTTP port (bound to 127.0.0.1)
CODEMIRA_POLL_INTERVAL_MINUTES	15	Minutes between daemon poll cycles
CODEMIRA_IDLE_THRESHOLD_MINUTES	60	Minutes before a session is considered idle
CODEMIRA_EXTRACTION_MODEL	z-ai/glm-5.1	Model name for memory extraction
CODEMIRA_EXTRACTION_BASE_URL	https://openrouter.ai/api/v1	OpenAI-compatible endpoint for extraction
CODEMIRA_EXTRACTION_API_KEY	—	Required. API key for the extraction endpoint.
CODEMIRA_SUBCORTICAL_MODEL	gemma4:e2b	Model name for subcortical / compression / link classification
CODEMIRA_SUBCORTICAL_BASE_URL	http://localhost:11434/v1	OpenAI-compatible endpoint for the subcortical roles
CODEMIRA_SUBCORTICAL_API_KEY	""	API key for the subcortical endpoint (empty for Ollama)
CODEMIRA_CONSOLIDATION_MODEL	gemma4:e4b	Model name for memory consolidation
CODEMIRA_CONSOLIDATION_BASE_URL	http://localhost:11434/v1	OpenAI-compatible endpoint for consolidation
CODEMIRA_CONSOLIDATION_API_KEY	""	API key for the consolidation endpoint
CODEMIRA_ARC_MODEL	gemma4:e2b	Model name for arc summarization
CODEMIRA_ARC_BASE_URL	http://localhost:11434/v1	OpenAI-compatible endpoint for arc generation
CODEMIRA_ARC_API_KEY	""	API key for the arc endpoint
CODEMIRA_MAX_SURFACED_MEMORIES	8	Max memories per retrieval
OPENCODE_DB	auto-discovered	Override path to OpenCode's global database

Memory stores are always located at <project-worktree>/.codememory/ — there is no configurable data directory and no global fallback.

Test

# Python tests
cd daemon
pip install -e .
cd .. && python -m pytest tests/ -v

# TypeScript tests
cd plugin && bun test

Note: Embedding tests require the MongoDB/mdbr-leaf-ir-asym model to be downloaded (first run will fetch it). Remote-extraction tests require CODEMIRA_EXTRACTION_API_KEY in the environment.

Development

For faster iteration during development, lower the poll and idle thresholds:

export CODEMIRA_EXTRACTION_API_KEY=sk-or-v1-...
CODEMIRA_POLL_INTERVAL_MINUTES=1 CODEMIRA_IDLE_THRESHOLD_MINUTES=1 python -m codemira.daemon

This makes the daemon check for idle sessions every minute and treat sessions as idle after just 1 minute of inactivity (defaults are 15 and 60 respectively).

Verify extraction with:

sqlite3 <project-worktree>/.codememory/memories.db \
  "SELECT id, category, substr(text,1,80) FROM memories WHERE is_archived=0 LIMIT 20;"

Verify retrieval by hitting the daemon directly:

curl -s -X POST http://localhost:9473/retrieve \
  -H "Content-Type: application/json" \
  -d '{"query_expansion":"your query","entities":[],"pinned_memory_ids":[],"project_root":"/path/to/project"}'

Health check:

curl -s http://localhost:9473/health

⚠️ Initial Implementation

This is an early release under active development. There will be bugs that tests didn't catch. The integration between the daemon, plugin, and live OpenCode sessions has not been exhaustively tested. Use with that in mind, and file issues when something breaks.