A narrated proposal for the team: what changed, why it matters, the trade-offs, and what we're asking for.
Three systems that turn daily operational noise into durable knowledge — automatically, every night. Nothing changes without human approval.
Every day, Warren logs directives from Tony, decisions with clients, dashboard rules from Joana, infrastructure changes, strategy shifts. Each day gets its own work journal. In a month, that's 50+ journals — thousands of lines where standing rules are buried alongside debug logs and one-time fixes.
Each solves a different part of the knowledge decay problem. All run at 03:00 PT every night. None alter canonical memory without human approval.
Reads daily journals, classifies each entry as durable or transient, proposes updates to MEMORY.md. Human reviews and approves before anything changes.
Compresses yesterday's raw journal into a dense checkpoint. Same file, fewer lines, better signal. Original preserved in git. Runs automatically after consolidation.
Scans what Warren kept searching for across sessions. Surfaces patterns nobody explicitly wrote down. Writes to its own diary — never alters MEMORY.md.
The systems activated today cost nothing extra. We're proposing one additional upgrade: better embedding model for search accuracy — enabled by the compression savings.
| Component | Before | After | Monthly Cost |
|---|---|---|---|
| Nightly Consolidation | Manual / never | Automatic, propose-only | $0 (uses existing cron) |
| Journal Compression | N/A | 1 API call/night | ~$0.15/night → $4.50/mo |
| Dreaming | Disabled | Enabled, conservative | $0 (gateway internal) |
| Embedding: 3-small → 3-large | $0.07/mo (776 chunks) | $0.26/mo (493 chunks) | +$0.19/mo |
| Total additional cost | — | — | ~$5/month |
Reads what was written. Proposes what to keep. Never decides alone.
Spawns an isolated session. Reads every daily journal since last distillation.
AutomaticDURABLE (cross-session fact) · TRANSIENT (resolved) · TOPIC (belongs in specialized file) · SUPERSEDES (updates existing)
AutomaticCreates proposed file + safety archive. Validates line count, pointers, idempotency. Does not touch MEMORY.md.
AutomaticOperator reviews line by line. Nothing promoted until explicit "approved."
Human approval requiredRaw journal → dense checkpoint in-place. Original in git. Search index re-processes automatically.
AutomaticReads what was repeatedly searched for. Surfaces patterns nobody explicitly wrote down.
Memory plugin scans the recall store — every search query and fact retrieval from the past 30 days.
AutomaticMust pass all gates: accessed ≥3 times, relevance ≥0.8, ≥3 unique queries. Recent facts weighted higher (14-day half-life).
Conservative defaultsPatterns go to memory/.dreams/. Does not alter MEMORY.md or daily files. Read-only output for operator review.
Each layer has different reliability. The lifecycle moves important facts up the stack.
MEMORY.md: 89 lines of standing rules, client registry, infrastructure facts, strategy. Loaded in every session, no search needed. This is the target.
Topic files (active-pending.md, pipeline-knowledge.md). Load when the right context triggers. Precise when triggered correctly.
Compressed daily journals. Searched via embeddings + keywords. Now 493 dense chunks instead of 776 noisy ones. Better, but still probabilistic.
If there are new durable facts, a proposed file waits for review. If nothing changed, nothing proposed. No noise.
Consolidation proposes, never promotes. MEMORY.md is never overwritten automatically. Any operator can approve.
The raw journal from yesterday becomes a 20-60 line checkpoint overnight. Less context noise, better model attention.
Over weeks, patterns emerge: facts Warren keeps reaching for. Readable anytime at memory/.dreams/.
Directives, rules, terminology, strategy — captured daily, then consolidated into permanent memory. "Noted" = written to disk.
Every compression and consolidation gets a git commit. Every promotion gets an archive. Nothing is lost.
Three items. Two are already active with zero cost. One requires a config change.
Nightly consolidation (propose-only mode) + journal compression + dreaming (conservative defaults). Cost: ~$4.50/month for compression API calls. Everything else: $0.
Switch from text-embedding-3-small to text-embedding-3-large. Higher accuracy on search. Cost delta: +$0.19/month. Requires config change + re-index (one-time, ~30 seconds).
For 73% less noise in context, 49% better search accuracy, automated knowledge consolidation, and standing rules that never get forgotten.