Yuba WangAI Operations

All case studies

Memory that survives the session

One shared, locally hosted memory store, connected to four AI tools and fed from my own notes by a local model.

Status
Running
Stack
  • Hindsight
  • Ollama
  • OpenClaw
  • Claude Code
  • Codex
  • Hermes
  • Obsidian
  • gitleaks
  • Podman

Problem

Each AI tool I use had its own memory. None of them could see what the others knew. I kept explaining the same context again.

My earlier memory layer also drifted. New project files stopped flowing into it. The agent then gave answers from old notes.

What I built

First, I repaired the old layer. I wrote a clear rule for when memory must be used. Every write had to be checked by reading it back, not by trusting a quiet “saved” signal.

Then I cleaned it up carefully. I measured where the noise came from. One area held over 7,000 stored entries (one snapshot, June 2026). It was a filing problem, not a disk problem. I added a barrier so that area stops filling up. I did not bulk-delete anything.

Then I built one shared store. I used Hindsight, self-hosted and pinned to a fixed version.

  • A local model on my own GPU pulls facts out of my notes. No hosted model is used for this step.
  • An hourly job feeds in new and changed notes.
  • Claude Code, Codex, Hermes and OpenClaw are all connected to it.
  • Agents can only read. The one writer is the ingest job, and it has gates.
  • Every file passes an allowlist, a denylist and a secret scanner before it is sent.

How it’s checked

  • The secret scanner has a 35-case test suite that runs through the real tool.
  • Counts are checked from two sides: what the store reports and what the ingest job recorded.
  • Cross-tool reading was tested live by comparing each tool’s document count with the store’s count.

Results

  • Backfill complete: 247 of 247 documents ingested. The store’s count matched the ingest record (October 2026).
  • Secret scanner: caught 6 of 6 known real credentials in testing (October 2026).
  • Live reading: Claude Code, Hermes and Codex each returned the same document count as the store when tested (October 2026).

What’s still open

  • Old layer not retired. The previous memory system still runs alongside. Its retirement is planned, not done.
  • Not every tool proven live. Claude Code, Hermes and Codex have been read from live. OpenClaw is connected but had not been tested in a live turn at the last status update.
  • Restarts. The store was down for about two days after a laptop reboot in October 2026. I added an auto-restart service. A crash test brought it back in about a minute. A full reboot test has not been done yet.
  • Speed. About one minute per note, because the GPU is shared. Fine for hourly updates.
  • Cleanup. Pruning of old entries has not been run. It waits for an approval gate.