Shared Memory

Shared Memory Framework

1. The problem

Coding agents already have good local memory. Markdown files, project notes and indexed documentation work well — as long as the knowledge stays inside that project.

However, when I opened the same agent in a different repository, it couldn't reuse one of its own architectural insights from the previous project — not because the model had forgotten, but because its memory was tied to the old workspace.

That led me to build Shared Memory. Instead of sharing conversations, it shares Architectural Decision Records (ADRs) with provenance, allowing important engineering knowledge to outlive individual projects and become available across projects — and, as a consequence, across different agents and models.

It's not a replacement for local agent memory. It's organizational memory — the ADR shelf of your engineering practice, with a lifecycle.

Two things make that more than a shared notebook. The first is the **lifecycle**: knowledge that crosses projects also outlives its own correctness, so the memory must manage the life of what it holds — retiring stale content by promoting fresh, refreshing its indexes, re-examining its conclusions when the evidence turns. Without that, six months in you have three conflicting truths and no way to ask which one stands.

The second is an honest division of labor with the language model. The ADR record-keeping needs no LLM at all: decisions, evidence, provenance and supersession are exact bookkeeping, and the system reproduces them deterministically. The model is employed only where it **earns its keep** — compressing a section's facts into an index card you can read for a fraction of the tokens, distilling clusters of tested decisions into the insight they demonstrate for a project domain, and rebuilding both when the lifecycle retires what they were made from. Turn the model off and you lose the synthesis; you never lose the record.

2. What you get

One memory, shared by every tool