An open-source, three-tier memory architecture that automatically degrades and archives old agent context. You get the benefits of llm-wiki, a vector DB, AND cold storage.
pip install git+https://github.com/jeremykamber/strata-memory.git Human memory is efficient because it knows how to forget. Strata applies the same principle. A new memory lands hot in the active stratum, settles into cooled as it ages, and eventually sinks to the archive as compact JSON. Every transition is deterministic; no LLM calls.
Workflow
Strata runs on the filesystem. Plain markdown files and algorithmic decay, no external services. Memories settle deeper over time and resurface when they prove useful again.
strata init
Created ./strata_data
Initialized active stratum
Initialized cooled stratum
Initialized archive stratum
Strata creates three layers: active, cooled, and archive. The Janitor moves data between them using file age and access count, with zero LLM calls.
strata add user/preferences.md "# User prefers dark mode"
Written to: active/user/preferences.md
New memories land in the active stratum as plain markdown. This is the agent's working memory: hot, fast, and fully editable.
strata search "user preferences"
Score Tier Source 1.50 ACTIVE user/preferences.md 0.50 MEDIUM user/dashboard.md 0.25 ARCHIVE Rate limit prefs... Search cascades across all three tiers. Active results rank highest, cooled appears next, and archived matches carry a flag for rehydration. Recent context surfaces before distant memories, just like the human brain.
strata promote
Promoted user/dashboard.md (accessed 4 times, threshold 3)
Moved back to active/user/dashboard.md
Memories don't just sink; they resurface. When the agent reads a cooled file enough times, it gets promoted back to the active stratum automatically - no separate command needed. The file proved useful again.
strata rehydrate a1b2c3d4e5f6 --target=active
Restored to 1st Stratum (active): user/rate-limits.md
Archived files are never gone for good. Reading an archived file by its original path auto-rehydrates it to active, or use `strata rehydrate` to choose the target tier yourself.
Agentic coding, agent memory, and building autonomous systems that learn how to forget. Architecture notes, tradeoffs, and what I learn along the way.
Follow on Substack