The idea it is testing is that 'delete after 7 days' and 'keep the last 100' both throw away the wrong things; here a decay score weighs how recently and how often a memory was used against how important it is, and touch_memory lets use extend its life. Promotion to Markdown long-term storage is the escape hatch for facts you never want to lose. It is explicitly a research artifact, so treat it as a place to study the idea rather than a product to depend on.
A temporal memory server: memories decay over time and strengthen with use, with a two-layer design where frequently used ones get promoted to permanent long-term storage.
- save_memory and search_memory store and retrieve; search_unified searches short-term and long-term together
- touch_memory reinforces a memory so it lasts longer; promote_memory moves a well-used one to permanent storage
- read_graph, create_relation and cluster_memories build and traverse a knowledge graph of entities and relations
- consolidate_memories and gc run the decay-and-cleanup cycle; open_memories opens specific entries
- A decay score combining recency, frequency and importance decides what stays, rather than a fixed expiry
- Local storage — human-readable JSONL by default, or SQLite; long-term memory as Obsidian-friendly Markdown
Python 3.10+. Short-term memory lives in JSONL or SQLite under a local config directory; long-term memory is Markdown with YAML frontmatter and wikilinks. All data stays on your machine, with no cloud service and no telemetry. The project is licensed AGPL-3.0 and describes itself as a research proof-of-concept, not a production product.
One command — uv tool install cortexgraph