Memory Layer
Memory Layer gives coding agents persistent, project-scoped memory. It captures decisions, context, and rationale from each session, curates them to remove duplicates and noise, and stores the results in a graph and vector database so future sessions can recall what matters and how it connects. Frequently used memories are re-validated against your code, so what agents recall stays accurate as the project moves. It works with Codex, Claude Code, OpenCode and the OpenAI, Voyage and Ollama embedding APIs.
From nothing to a running stack with a demo project in about five minutes, with one command and no API keys.
Install Memory Layer, configure storage, and initialise your first project.
Review built-in loop workflows for memory hygiene, CI triage, context packs, PR review, and evaluations.
Copy a prompt and let your coding agent install and configure everything.
Memories earn trust through use and are validated against your code when they run hot.
Understand memories, curation, evidence, retrieval, and the code graph.
Search real memories and explore the graph in your browser — no install, no backend, nothing leaves the page.
Terminal UI
Browse, search, and inspect every stored memory from the terminal. The TUI shows metadata, confidence scores, embeddings, canonical text, tags, and the original source — so you can see exactly what your agents remember and where it came from and what it relates with.
Web UI
The service also serves a browser interface for dense inspection of memories, query evidence, activities, watchers, embeddings, errors, and replacement proposals. It uses the same backend APIs as the TUI, so it is useful when you want wider panes, screenshots, or a visual review surface while keeping the service local.
Open the self-contained demo to click through a static snapshot without installing or running a backend.
Automations
Memory Layer includes built-in loop automations for context-pack refresh, memory hygiene, CI failure triage, issue triage, draft PR preparation, reviewer drift detection, skill mining, and memory evaluation. They run through the local service control plane and produce reports, approval requests, or reviewable proposals rather than bypassing Memory APIs.
Self-maintaining memory
Every retrieval strengthens the memories that earned it; activation spreads to linked knowledge and fades with time. When a memory runs hot, Memory Layer gathers evidence from your repository and validates it: confirming it, improving unclear wording, or queueing a correction for review. Weak evidence never rewrites anything — it flags the memory instead. Inspect it all with memory scores, memory validate, and memory review.
Get set up by your agent
Paste the install and repo-init prompts into Claude Code, Codex, or any coding agent and let it install and configure everything for you.