What it is
RepoLens scans a repository into structured packs,
exposes that model through an MCP server, and pairs it with a
typed, decaying memory graph. The goal is not to write code for
the agent. The goal is to stop every new session from starting
cold.
Current private snapshot
Private pre-release snapshot: origin/main @ afd7a6b.
That private snapshot includes repo scanning, 26 MCP tools, tiered
summaries, convention extraction, pack comparison, grounded
long-term memory recall, repolens init,
brief, remember,
eval list/start, and warnings-only
validate-plan.
What's different about it
RepoLens treats repository understanding as structured perception, not as a bigger prompt dump. Memories have types, anchors, decay, and grounding status; tool responses carry epistemic metadata so an agent can see whether it is reading observation, heuristic, or interpretation.
Next
The planned v0.1 direction is agent legibility and memory safety: richer memory types such as hypotheses, constraints, failed attempts, and friction findings, plus explicit confidence labels. Those pieces should be described as planned until they appear in a public release or public main snapshot.
What it isn't
- Not an autonomous coding agent.
- Not a replacement for tests or human review.
- Not a generic RAG store.
- Not multi-repo or org-wide memory today.
- Not a guarantee that recalled memory is correct.