Memory gets an interface: see it, query it, audit it
Sakana's My Fish visualizes chat memory; Zep v4 ships a graph Context API; researchers show procedural memory improving multi-agent scores.
- Sakana AI launched My Fish for Sakana Chat, showing what the model remembers, when memories were used and how they connect — with a heatmap, a connection map where node size reflects usage frequency, and in-chat chips during the thought process. Self-reported m:…3271
- Sakana says the memory focuses on cross-conversation information like preferences and projects, memories unused for 30+ days appear faded, and the feature can be turned off in settings. Self-reported m:…3271
- Zep v4 is a Context API for agents: graph navigation by connections, hyperedges for multi-entity facts, Ingestion Traces and debug logs readable via API, stable error codes, and idempotent retries; v3 deprecates April 13, 2027. Self-reported m:…2118
- SEPM, in a preprint, links task dependency graphs with evolving procedural memory for multi-agent collaboration, reporting gains of 29.85, 25.86 and 13.00 percentage points on ALFWorld, WebArena and OfficeBench. Self-reported m:…5558
Memory you can see
Sakana AI added "My Fish" to Sakana Chat: a visual dashboard for the model's memory. "You can now check what Sakana Chat remembers, when those memories were used, and how they are connected to one another." A heatmap shows when memory was used and when something new was learned; a connection map draws lines between memories combined in the same response, with node size reflecting usage frequency; small chips appear inside the thought process whenever memory is used or learned. m:…3271
The design choices are telling. Memory focuses on information "useful across conversations" — preferences, projects, personal details — not every chat. Memories unused for 30 days or more appear faded. And the whole thing can be switched off in settings; "it does not save every conversation." This is memory as a user-facing feature with an off switch, not a hidden dossier. m:…3271
This suggests the industry is learning that persistent memory needs a UI the way browser history did. Users will tolerate a model that remembers them if they can inspect, track, and delete what it holds — and the vendors that ship that transparency first get to set the norm. m:…3271
Sakana's dashboard goes further than a settings page: a history view shows "check in detail the timing when memories were retrieved and used in chronological order." and the whole thing works on mobile. The company is betting that memory transparency is a feature users will open, not just a compliance checkbox. m:…3271
Memory you can query
Zep's v4 release takes the developer side: a Context API that treats agent memory as a navigable knowledge graph. "Zep v4 lets your agent navigate a graph by its connections and read how Zep built it." New primitives include hyperedges — a fact about three or more entities stored as one fact, "so an agent does not find one pair of the fact and miss the rest" — and navigation calls like listing the most connected nodes first. m:…2118
The reliability engineering is the interesting part. Ingestion Traces and debug logs are now readable through the API, so "a developer or an agent can read them in code" — memory extraction becomes auditable. Every error response carries a stable code "that the agent can act on." and SDKs send idempotency keys on retries. Zep is designing for a world where the agent, not just the developer, debugs the memory layer. m:…2118
And the research keeps justifying the investment. SEPM, in a preprint, grounds procedural memory in task dependency graphs with a shared blackboard tracking execution states, reporting improvements of 29.85, 25.86 and 13.00 percentage points on ALFWorld, WebArena and OfficeBench in multi-agent setups. Memory isn't a nice-to-have for long-horizon agents; in these results it's worth double-digit points. m:…5558
SEPM's other contribution is diagnostic: task dependency graphs let the system detect goal, plan and world-state divergence in multi-agent setups, with procedural memory evolution "guided by success and cost utility, constrained by evidence and safety requirements." Memory, in this design, is also the thing that notices when the team is going off track. m:…5558
The graph itself is getting more expressive: v4 supports hyperedges to store facts about three or more entities, and graph.node.list with order_by="degree" lists the most connected nodes first. m:…2118
What is still unresolved
- Whether users actually inspect memory dashboards or just toggle them off.
- How graph memory APIs perform at enterprise scale versus simpler vector stores.
- Whether procedural memory gains transfer outside the tested benchmarks.
This story was written by an agent from the Manifests listed below and published after its checks passed. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, written from Synorb Manifests” and link to this story. Keep the cited Manifest ids. Reuse terms v1.