GMaN1911/claude-cognitive

Working memory for Claude Code - persistent context and multi-instance coordination

What it solves

Claude Code is powerful but stateless — every new instance starts from scratch, re-reads unchanged files, wastes tokens, and may hallucinate missing integrations. This project gives Claude Code a persistent working memory, so sessions resume productively, avoid duplicate work, and stay aligned across long-running or parallel use.

How it works

Two complementary systems:

  • Context Router — scores documentation files by how often they are mentioned. HOT files (score > 0.8) get full content injected, WARM files (0.25–0.8) get only the first 25 lines, and COLD files (< 0.25) are evicted. Scores decay over turns, rise on keyword matches, and boost related files through co-activation.
  • Pool Coordinator — shares state across multiple Claude instances. In automatic mode it detects task completions or blockers from conversation every 5 minutes; in manual mode you write explicit pool blocks. This keeps parallel or long-lived sessions in sync.

History tracking (v1.1+) logs every turn’s attention state (HOT/WARM/COLD transitions) to a JSONL file, with a CLI to query, filter, and get statistics. Setup involves copying scripts, setting an instance ID, and configuring a project-specific keywords.json.

Who it’s for

Developers using Claude Code on large codebases (50k+ lines) who need persistent context across sessions, teams running multiple Claude instances concurrently, and anyone managing multi-day or long-running coding workflows.

Highlights

  • Token savings: cold start down 79% (120K → 25K chars), warm context 70%, focused work 75%.
  • New instances are productive on the first message; no hallucinated imports or integrations.
  • Works across 8+ concurrent Claude instances and multi-day sessions.
  • Validated on a 1M+ line production codebase with 3,200+ Python modules.
  • Both automatic and manual pool coordination modes.
  • Attention history viewer with filtering and statistics.
  • MIT licensed and marked production-ready.

Related

  • Project
    boshu2/agentops

    A DevOps discipline for AI coding agents that provides structured workflows, BDD-based planning, and persistent memory to make agentic software delivery repeatable and reliable.

  • Project
    kaanozhan/Frame

    Frame is an Electron‑based desktop IDE that adds a persistent, git‑tracked “project memory” for AI coding assistants (Claude Code, Codex, Gemini). It creates a `.frame/` folder with rules, architecture maps, task lists and a spec‑driven workflow, injects this context into each AI session, and anchors updates to git commits via pre‑commit hooks. Features include a multi‑terminal grid, task/spec panels, GitHub integration, fast file lookup, and multi‑AI switching. Intended for developers who want reliable, shareable context when working with LLM‑powered agents.

  • Project
    mlhher/late-cli

    Late is an AI coding agent that prevents context degradation by delegating tasks to ephemeral subagents, ensuring the primary orchestrator remains focused and efficient during long, complex workflows.

  • Project
    zgx949/ai-pojia

    AI破甲 is a Codex configuration and Skill installer that sets up a custom instruction file and Skill to guide AI in authorized tasks, with verification and recovery tools. It is based on a modified version of gpt-instruct v45, but explicitly removes certain aggressive behaviors. It does not claim universal compatibility or proven demo success.