Our Memory
Recall Experiment
How nature inspired a little memory experiment.
A few caveats
- We’re still experimenting. Nothing here is final — we’re learning as we go and changing our minds often.
- There are many open-source projects tackling this space. We looked at a lot of them. OpenViking (
github.com/volcengine/OpenViking) is a great starting point if you want to explore this area. - Our philosophy: harness first, infra later. Shape the knowledge workflows bottom-up from our actual needs, then figure out the infrastructure and tooling once we know what we’re building for.
- This is not a source of truth. It’s a snapshot of something we find interesting. Take it as a conversation starter, not a recommendation.
In one breath.
If you only read three things from this deck — read these.
- 01 · The problemKnowledge dies at two boundaries. Between sessions (time) — intent disappears when the transcript ends. Between agents (space) — reasoning stays locked in individual transcripts. We lose what was learned in the moment.
- 02 · The stackThree layers, one brain. Fully local. Claude Code as the harness, PRISM as the knowledge cycle, AuraSDK as the memory store. Knowledge persists as a side effect of working — no second system to feed.
- 03 · What we foundTwo claims worth pressure-testing. Emergence is a primitive-design problem, not a package feature. And memory is a four-quadrant substrate — humans and agents both produce, both consume.
The Problem
Two Problems: Time + Space
When a session ends, the intent that shaped the artifacts disappears. When parallel agents work, their reasoning never leaves their transcripts. Knowledge dies at two boundaries.
A user says: “I want it to feel like @mentions”. That single sentence drives the concept doc, three ADRs, the spec. When the session ends, the artifacts survive but the intent disappears.
- Next session: the agent has the spec but doesn’t know why it says what it says
- Corrections vanish too — “don’t use that plugin”, “not that folder name”
An orchestrator dispatches a researcher, an analyst, a concept builder. Each produces an artifact, but the reasoning and dead ends stay locked in their individual transcripts.
- Agent A’s abandoned approach might be exactly what Agent B needs
- The orchestrator sees outcomes — not the journey, not the institutional knowledge
The Stack
Harness First, Infra Later
The hard part isn’t the code. It’s the knowledge in the developer’s head — decisions, constraints, preferences, corrections — that never makes it into the codebase. Agents can read code. They can’t read intent.
ETH Zurich’s Evaluating AGENTS.md (Feb 2026) found that LLM-generated context files degraded task success by ~3% while increasing inference cost over 20% — they pre-cache what the agent would discover on its own.
Human-written context improved performance only ~4% on average, and only when minimal and scoped — i.e. focused on non-inferable knowledge: the constraints, judgments, and preferences that can’t be derived from reading the code.
Source: Evaluating AGENTS.md, Gloaguen et al., Feb 2026
We pick infrastructure that captures non-inferable knowledge as a side effect of working. Then we build the workflows that surface it. Three layers, in order: