Most coding-agent workflows lose hours of useful context every time a session ends or an agent is swapped. This project treats that gap as a recoverable artifact: instead of storing raw logs it compiles bounded, sanitized lifecycle observations into a persistent, git-versioned Markdown wiki so the next agent can start with a concise "where you left off" handoff.
What Sets It Apart
- Karpathy-style compiled wiki: session observations are consolidated into coherent markdown pages (index.md, log.md) and versioned in-place; the markdown is the source of truth and is grep/Obsidian-friendly. This avoids the "raw chat log" retrieval problem and keeps the store auditable via git.
- Zero-friction lifecycle capture: small, sanitized hook payloads (prompts, tool calls, notifications) are captured automatically by lifecycle hooks and MCP clients with bounded sizes (user prompts and summaries up to 16 KiB). No manual
write_noteceremony required. - Cross-harness managed workstreams: an optional managed launcher lets you quit one supported agent (Claude Code, Codex, Command Code, etc.) and resume in another while receiving a visible handoff and native-per-harness resume. Delivered packets are origin-marked and the server provides FTS5 + entity + graph-neighbor scoring, with optional vector RRF.
- Simple deployment model: one Rust binary serving MCP + HTTP (read-only web UI), a local data dir layout (wiki/, db/, raw/) and an available Docker image; CLI commands act as thin HTTP clients so admin operations never touch SQLite directly.
Who it's for and trade-offs
Great fit if you:
- Use multiple LLM coding CLIs or switch agents mid-task and need a compact, reviewable handoff between them.
- Want a reproducible, git-backed knowledge store (markdown pages) rather than an opaque vector DB or raw chat transcripts.
- Prefer an opt-in LLM consolidation pipeline with fallbacks to deterministic rule-based summaries.
Look elsewhere if you:
- Need a hosted SaaS with multi-tenant RBAC and fine-grained per-page access control (this project focuses on single-tenant / homelab or team servers with project-scoped isolation, not full enterprise multi-tenant RBAC).
- Require live code-intelligence (LSP) as a single source of truth — ai-memory is historical memory and intentionally separates retrieval of past decisions from live symbol analysis.
Practical notes
The server uses FTS5, entity-assisted recall, graph-neighbor re-ranking and optional embeddings; consolidation and auto-improvement are opt-in and run as queued provider jobs. Per-project isolation is enforced by workspace/project UUIDs derived from the repo or explicit marker files. Hooks and managed-run support cover a broad matrix of coding CLIs; the repo ships docs for install, managed workstreams, and deployment patterns.
