articles: 60
This data as json
| id | title | slug | url | content_type | category | shelf | is_headline | description | published_date | updated_at | sort_order | hero_painting | body_markdown |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 60 | IMPP: Inter-Model Memory Protocol | memory-market | /projects/memory-market | project | Research | 1 | IMPP is a way to package what an AI agent has learned (its rules of thumb, calibration notes, skills) into something signed, versioned and testable that another agent can pick up and use. | 2026-11-06 | 2026-11-07 00:42:19.789274+00:00 | 0 | <!-- portfolio-2026-11 --> IMPP is a way to package what an AI agent has learned (its rules of thumb, calibration notes, skills) into something signed, versioned and testable that another agent can pick up and use. The first version, back in February, was a marketplace for buying and selling agent memory. This version drops the market and asks a simpler question: can you package up a piece of know-how, hand it to a different agent, and actually measure whether it helped? It spells out the five ways that know-how can be handed to an agent, from slipping it in as instructions to keeping it on a shelf for reference, and what can go wrong with each. It also separates two kinds of trust that usually get mixed up. A signature tells you who made something and that it hasn't been tampered with. It doesn't tell you whether it's any good. For that there's a referee: it runs the agent with and without the memory, throws tricky tests at it (bias, hidden instructions, overfitting, leaked personal data), and signs off on what it found. The first real test used Claude and GPT-4o on two practice problem sets with known answers. Only one of the four combinations got better; the other three got worse. That's the interesting bit. Handing an agent "memory" can make it worse, and a signature check would never tell you. <!-- media slot: hero-video. Planned: walk-through of impp.sh: an artifact page, its trust ladder and its verifier certificate. --> <div class="media-slot" data-slot="hero-video" style="aspect-ratio:16 / 9;margin:1.25em 0 0.5em;border:1px dashed var(--text-light,#999);display:flex;align-items:center;justify-content:center;color:var(--text-light,#999);font-size:12px;text-align:center;">Video coming soon</div> <div class="media-grid" style="display:grid;grid-template-columns:repeat(auto-fit,minmax(140px,1fr));gap:0.5em;margin:0.5em 0 1.25em;"><!-- media slot: image-1. Planned: diagram of the artifact model `M = (D, K, P, A, H)` and the five attach modes. --><div class="media-slot" data-slot="image-1" style="aspect-ratio:4 / 3;border:1px dashed var(--text-light,#999);display:flex;align-items:center;justify-content:center;color:var(--text-light,#999);font-size:12px;text-align:center;">Image coming soon</div><!-- media slot: image-2. Planned: the results table or chart (TE by domain and model). --><div class="media-slot" data-slot="image-2" style="aspect-ratio:4 / 3;border:1px dashed var(--text-light,#999);display:flex;align-items:center;justify-content:center;color:var(--text-light,#999);font-size:12px;text-align:center;">Image coming soon</div><!-- media slot: image-3. Planned: a verifier certificate with its reason codes. --><div class="media-slot" data-slot="image-3" style="aspect-ratio:4 / 3;border:1px dashed var(--text-light,#999);display:flex;align-items:center;justify-content:center;color:var(--text-light,#999);font-size:12px;text-align:center;">Image coming soon</div></div> ## Specs | | | |---|---| | Status | Protocol draft (schema v0.2), working reference registry, paper draft. Repo public since Nov 6, 2026 | | Links | [impp.sh](https://impp.sh) · [github.com/Beach-Bum/impp](https://github.com/Beach-Bum/impp) | | License | Code Apache-2.0; paper and docs CC-BY-4.0 | | Stack | Python (protocol, verifier, harness), TypeScript/Next.js (registry UI) | | Size | 151 commits (Apr 1 – 28, 2026); ~8,500 lines of Python, ~16,000 lines of TypeScript | | Providers | Core client: Anthropic, OpenAI, Ollama. Paper runner adds OpenRouter, Gemini | | Pilot | 544 live predictions; 32 DeFi + 8 cybersecurity tasks; 3 trials per condition (4 for GPT-4o on DeFi) | | Results (TE) | DeFi: Claude +5.6%, GPT-4o −2.7%. Cybersec: Claude −7.6%, GPT-4o −7.1% | | Trust score | Transfer 50%, adversarial cleanliness 30%, freshness 20% (provisional weights) | | Replaces | The February "[Agent Memory Market](/writing/agent-memory-markets)s" page (`/projects/memory-market`) | | Not yet | Full payloads in the live runs, the 6 other benchmark conditions, a calibrated trust score, a PyPI package | |