LLM Memory Tool Memory
Information, Legal, Security

Memory Infrastructure For LLMs And Agent Workflows

Turn large chats, files, notes, logs, and project folders into private searchable memory for AI assistants.

The app stores compressed local chunks plus small search headers, so assistants load only the evidence they need.

Use it with Claude, ChatGPT, Codex, or local/Ollama workflows when the chat or file is available on this machine.

This website is only for user information, legal/security references, and the official download placeholder.

Download Placeholder (.exe)

Version: alpha · Size: TBD · SHA256: checksum file

What This App Does

  • Persistent memory for long chats/projects.
  • Retrieval over replay (token savings).
  • Better continuity and less drift.
  • Ingest documents for planning and research.
  • Cross-reasoning between multiple files.

Why It Helps Reasoning

  • compresses huge sources into local chunks
  • searches headers before loading raw text
  • reduces token use and long-chat drift
  • tested with a 96.8M-character chat

⚡ Workflow and Research Use

Codex Baseline

Deterministic long-chat workflow (Codex baseline):

  1. Start from distilled summary/policy, not full raw history.
  2. Retrieve narrow first (limit 3-8), then hydrate only evidence needed now.
  3. After every 4th compression cycle, run checkpoint:
    - retrieve distilled summary + policy
    - reason over newest raw segment (evidence-first)
    - ingest raw segment compressed (hydrable)
    - write one distilled summary with:
    facts_verified, decisions, guardrails, mistakes_and_corrections, open_questions
  4. Supersede prior summary explicitly, keep one active distilled summary.
  5. If retrieval quality drops, run checkpoint immediately.
  6. Other LLMs compress at different stages: tune checkpoint timing + prompts per client/model.

Per-Model Adaptation

Use separate workflow tuning per client/model:

  • Claude / ChatGPT / Ollama may compress at different stages.
  • Keep checkpoint trigger and prompt template per model.
  • Validate with retrieval quality checks before scaling.
  • Do not assume Codex cadence transfers unchanged.

Website Scope

Support: deepvalueagent@gmail.com

(c) Martin Freund · Last updated: 2026-05-03