Three feeders → one memory folder

Every agent in the office has a small memory folder (~10-20 .md files) that loads at every cold-start. Three streams feed it.

The three feeders

Feeder	What it gives	Source
Office library	Stories built from emails, PDFs, photos, ERP rows — through 5 perspectives	The user’s computer + cloud connectors
Chat reflection	The agent’s own working thoughts from past conversations	Claude Code JSONL session files
Pinned facts	Atemporal rules the agent must always know	Hand-curated, lasting truths

All three converge into the memory folder, which loads at every cold-start (~10K tokens of “where am I in the world right now”).

Why three feeders, not one

Each feeder answers a different question the agent has at wake-up:

• Office library answers: “What happened in the last 24 hours that I should know?”
• Chat reflection answers: “What did I say last time about this topic?”
• Pinned facts answers: “What is always true that I must not forget?”

A single feeder collapses these into noise. Three lets the agent reason cleanly about why a piece of context exists.

The cold-start inject

At SessionStart, the agent’s hook reads from all three:

[Tier 1 inject, ~10K tokens]
├── Final story (today's body — from office library)
├── Top 5 active casefiles (from office library)
├── Pinned facts (5-10 lines, hand-curated)
├── Recent notes (last 24h, from chat reflection)
├── Leave note (what I was doing last time)
└── Last 20 Q→A pairs (from chat reflection)

Two more tiers are available on demand:

• Tier 2 — skill API for mid-conversation queries (recall.py --topic X)
• Tier 3 — provenance walk back to raw bytes when audit needed

Read next

Multi-perspective How office library files become 5 post-its through 5 lenses.

Provenance The chain from final story → story → post-it → evidence → raw bytes.

Worked example · One AI incident The three feeders in action — one incident walked through all five stages.