Hello, colleagues. I've bitten off something ambitious. Analyzing the problems surfacing on my own project ShotPorter https://ShotPorter.com, I decided I needed a long-term memory system. Looking at what already exists, I found nothing that solved my problems, so I built my own. After using it for a while, I decided it was worth polishing and releasing publicly. It became MemContinuum — long-term decision memory for Claude Code projects. And it is not just another memory system, collecting facts happened and words written. **The problems it addresses.** On a long project you forget what was decided about a given question and why. The model forgets harder. Subagents know nothing at all — the orchestrator dispatches them nearly blind onto narrow tasks. The result is reinvention instead of reuse: duplicate implementations, drift, tokens burned re-solving solved problems, and settled questions resurfacing as "wait, why is this written this way?" **How it differs from the memory systems I looked at:** - Two layers: ANATOMY - an indexed map of the code, and RATIONALES - decision chains recorded against it — what was decided, **who** decided it, how it changed over time, plus incidents and rejected alternatives with reasons. - Reading is automatic. Before an agent edits a file, the decision chain governing that path is injected into its prompt. Nobody has to remember to look. - Writing is unavoidable but not automatic. The agent gets a question it must answer; "nothing to record" is a legitimate answer. It's moderated by judgment, not a scraper dumping everything into a pile by keyword or timestamp. - One AI writes records — the orchestrator. Subagents and external reviewers (Codex, Grok) propose through an inbox; proposals become records after triage. - Per project, local, no server. Markdown as the source of truth, SQLite as a disposable index. Built for coding projects specifically: without indexable code only half the brain works. Current state: 0.2.0rc4, honestly labelled a release candidate. MIT. Claude Code only for now. The README is long and detailed if you want the full picture. Feedback of any kind is very welcome. Besides me, a team of authors worked on this project: - Claude Code: Fable 5/5.1 as lead engineer and project manager; Opus as inspector; Sonnet as coder; Haiku as tester - Codex: 5.6 Sol / 6 Astra as reviewer and outside consultant - Grok 4.6 as second reviewer https://github.com/krakozavr/Mem...

MemContinuum — memory for Claude Code

MemContinuim is linking chain of decisions with the codebase

MemContinuim is linking chain of decisions with the codebase