FIND-20260323-029 — claude-subconscious — Persistent Cross-Session Memory for Claude Code Agents

claude-subconscious is an experimental TypeScript library by Letta AI (creators of MemGPT) that injects persistent memory into Claude Code sessions via four hooks: SessionStart, UserPromptSubmit, PreToolUse, and Stop. A background Letta agent watches each session, reads the codebase, accumulates memory over time using a git-backed filesystem (memfs), and whispers context back before each prompt — all without modifying files in the working directory. It supports multi-project shared memory, three verbosity modes (whisper/full/off), and SDK tool access (Read, Grep, Glob, web search). The project is explicitly experimental (v0.x), MIT-licensed, 1044 stars, actively maintained with the last commit on 2026-03-18.

Directly addresses the ADLC pipeline's current state-loss problem: agents lose context between dispatcher-v3.sh cycles every 5 minutes. The hook-based whisper pattern maps precisely to the ODS agent lifecycle — SessionStart could initialize pipeline state, Stop could serialize agent decisions to the Letta memory store, and UserPromptSubmit could inject relevant service context before each dispatcher command. The git-backed memfs approach (last commit: migration to memfs) aligns with ODS's file-based agent-memory system at ~/.claude/agent-memory/. Multi-project support is directly relevant given the ODS multi-service registry. However, the hook system requires Claude Code hooks, which adds a dependency on Claude Code's settings.json injection mechanism — the same mechanism flagged in CVE-2025-59536 and CVE-2026-21852 (malicious repo settings overriding user approval). Use with caution in the ODS context: validate hook injection is controlled and not exposed to untrusted repository configs.

claude-subconscious — Persistent Cross-Session Memory for Claude Code Agents

Source

ODS Impact

Security Review

Tags