ctx-memory AGENTS.md

Project Memory (from previous sessions)

Project Memory

Architecture

• Persistent memory layer for LLM coding agents (Claude Code, Codex, Gemini, OpenCode) backed by SQLite + sqlite-vec, exposed via MCP
• Three-layer pipeline: Layer 1 (pure message/event processing) → Layer 2 (digest ≤500 tokens) → Layer 3 (project memory markdown merge)
• Shell wrapper intercepts tool binaries, spawns MCP sidecar (Claude/OpenCode only), runs pipeline on session end
• Claude: hooks fire live (PostToolUse) + JSONL reader seeds messages post-session
• Codex: post-hoc only — reads ~/.codex/state_5.sqlite + rollout JSONL after session ends
• OpenCode: MCP sidecar + AGENTS.md injection at startup
• AGENTS.md used for context injection for Codex and OpenCode (written before spawning tool)
• Non-interactive subcommands (update, mcp, login, etc.) bypass wrapper entirely via execFileSync
• You want OpenClaw (from Discord) to be able to act directly on your local files instead of only replying with paste-ready output. I’m checking your local OpenClaw/GSD setup to see exactly which permission boundary is blocking that and what to change safely.
• I successfully applied the config-side changes, but openclaw approvals set failed because its CLI syntax is different than expected. I’m switching to directly updating ~/.openclaw/exec-approvals.json, then I’ll verify and restart gateway.
• Now I have the full picture. Here are all the problems:

---

• Problems Found

• 1. Project12 sessions go to wrong project ("ghadi")

• Root cause: /home/ghadi is a git repo (has .git/). When you run claude from /home/ghadi/perso/Project12, git rev-parse --show-toplevel walks up and returns /home/ghadi — so the wrapper resolves it to the ghadi project instead of creating a Project12 project.
• Fix needed in src/wrapper/project.ts: If gitRoot is a parent of cwd (not the same or a direct project root), fall back to using cwd itself.

• 2. Architecture/decisions never saved in digests

• Root cause: decisions in Layer 1 only come from:
• • "confirmation" pattern matches with extractedDecision
• • Messages typed as "decision"
• When Claude announces architecture, it's just an assistant message — Layer 1 never classifies it as a "decision". So decisions: [] always, nothing reaches Layer 3's architecture section.
• Fix needed: Layer 1 needs to extract decisions from assistant messages that match architecture/decision patterns (e.g. "I'll use...", "The architecture is...", "decided to...").

• 3. llm-memory projects list didn't show Project12

• Root cause: Same as #1 — Project12 was never registered. Its sessions went to the ghadi project.

• 4. No messages captured (0 messages_raw despite JSONL existing)

• Root cause: The JSONL reader runs at session end and finds the file by mtime >= sessionStartMs. But the current session's JSONL file is still open/being written, so the mtime check may race. Also, 0 messages in DB means seedBuffers with messages never ran (no messages seeded from JSONL).

---

• Want me to fix these now? The priority order I'd suggest:
• 1. Fix #1 (git root leaking to parent) — affects project isolation completely
• 1. Fix #2 (decisions extraction) — affects architecture/memory quality
• 1. Fix #4 (messages not seeded) — investigate why JSONL reader returns nothing for recent sessions
• Now fix tryConfirmation in patterns.ts to extract a meaningful decision sentence instead of dumping the whole assistan…
• "I'm using" still matches "using " in ARCH_SIGNALS. And "React 19 instead of 18" matches "react". Both are still appeari…
• Clean! But "I'm using Subagent-Driven Development to build KubeQuest task by task" is in Architecture because it matches…
• The real fix is to change the entrypoint guard to use import.meta.url instead of VITEST, but actually looking at th…
• The fix: check if argv[1] ends with wrapper/index.js (the only valid case to run main()), rather than checking fo…