1. Markdown memory
Your workspace knowledge lives as local Markdown: context, observations, runbooks, decisions, actions, and session snapshots.
Local-first memory gateway for Pi
Stop paying the repo rediscovery tax.
pi-memctx gives Pi a durable Markdown memory layer: it recalls runbooks, architecture decisions, project context, and prior discoveries before the model starts burning tools and tokens.
66.9%faster in real-world stress run
95.0%fewer tool calls
80.8%fewer provider tokens
+25scored quality facts
Benchmark: same prompts, less rediscovery.
Real-world stress benchmark: 5 anonymized memory packs, 5 repeats per task, plain Pi baseline vs Pi with pi-memctx gateway. Names are anonymized as Pack 1–5; raw private pack data is not published.
| Profile | Runs | Avg latency | Provider tokens/task | Tool calls/task | Quality | Timeouts |
|---|---|---|---|---|---|---|
| Baseline | 25 | 196.6s | 34,385 | 27.2 | 206/255 | 1 |
| Gateway | 25 | 65.0s | 6,614 | 1.36 | 231/255 | 0 |
| Public synthetic release check | Baseline | Gateway | Delta |
|---|---|---|---|
| Average latency | 22.0s | 10.5s | 52.2% faster |
| Tool calls/task | 6.2 | 0.3 | 95.2% fewer |
| Visible tokens/task | 571 | 329 | 42.4% fewer |
| Quality | 63/110 | 76/110 | +13 facts |
How it works
2. Gateway retrieval
Before Pi answers, the extension detects the active pack, searches relevant notes with qmd or grep fallback, and injects a compact brief.
3. Bounded fallback
If memory is partial, Pi can still inspect source files, but with a budget to avoid runaway rediscovery.
Local. Inspectable. Open source.
No hosted memory vendor. No hidden database. No domain-specific hardcode. pi-memctx is designed for many languages, countries, teams, and stacks.