# Tagged MemoryAgentBench

Research and lab write-ups from drksci tagged MemoryAgentBench, newest first.

- [Engram](/engram): A memory should change when you read it as well as when you write it. Ten ways to try, and one test they all have to pass.
- [Bloodhound](/engram/bloodhound): Bloodhound trains a very small model to choose the next step in a walk through memory, using paths read backwards from known answers.
- [Canopy](/engram/canopy): Canopy hands the organising of a memory to a language model working offline: it groups entries into topics and writes a short summary at each level.
- [Coral](/engram/coral): The headline was 93%. It was fourteen of fifteen questions on a thirty-five-fact set the author wrote.
- [Glyph](/engram/glyph): Glyph stores memory in a compact notation, and the first lines of the store define the notation.
- [Horizon](/engram/horizon): Horizon places memories in hyperbolic space, where there is room for a tree: general memories near the middle and specific ones out towards the edge.
- [Loom](/engram/loom): Loom settles a conflict by how much of the rest of memory agrees with each side, not by the date.
- [Pheromone](/engram/pheromone): Pheromone lets a read leave a trail along the paths that worked, so later reads walk where earlier ones did.
- [Ring](/engram/ring): Ring holds candidates in a small working set and makes them survive repeated re-scoring against the query and each other before any of them is returned.
- [Tide](/engram/tide): Tide is the control for every claim about conflict resolution.
- [Web](/engram/web): The fact you need often shares no words with the question. It shares a name with a fact that does.