CodeMira: Developer Memory System

2026-05-14 · source · source ↗ · raw capture · wiki page

Proof-of-concept adapting Mira OS memory architecture for programming contexts. Learns from coding sessions and surfaces relevant context when needed.

Architecture

Python Daemon: - Monitors OpenCode sessions during idle periods - Extracts patterns from tool-call transcripts via LLM - Stores in SQLite with vector embeddings (hnswlib) + full-text search (FTS5) - Serves retrieval over HTTP

TypeScript Plugin: - Integrates with OpenCode's message transformation pipeline - Analyzes conversation intent via LLM - Queries daemon for relevant memories - Injects context into message stream

Features

Project-Scoped Storage: .codememory/memories.db per project
Hybrid Retrieval: BM25 + vector similarity via ANN + Reciprocal Rank Fusion
Multi-Provider: Ollama, OpenRouter, vLLM, llama.cpp, OpenAI-compatible
Deduplication: Fuzzy matching + vector similarity

Origin

"Concepts are adapted from Mira and not ported 1:1 as a tool-call laden coding session is different than an ongoing conversation."

Early release, active development.