In Part 1 I explored why durable, explainable episodic memory is essential for agents that need to operate reliably over long horizons. The argument was simple: an agent’s memory must be more than a place to store fragments. It needs to be inspectable, testable, and able to show how the past shaped the present.
With Mnemara v0.5.0, I’ve shipped three capabilities that directly strengthen that foundation of trust: judged recall evaluation, time-travel recall, and append-only changefeeds. Together, they move Mnemara from remembering well towards proving, replaying, and observing how memory behaves over time.
Judged Recall Evaluation Harness
One of the biggest gaps in most memory systems is the inability to measure whether recall is actually good. v0.5.0 introduces a structured evaluation harness for judged recall.
You can now define assertions of four kinds:
- Expected - items that must appear in the result set
- Optional - items that are acceptable but not required
- Disallowed - items that must not appear
- Explanation notes - free-form context for human or LLM judges
This turns recall from a black box into something you can systematically test, score, and improve. It’s a foundational piece for anyone building production agent systems where you need confidence, repeatability, and auditability in what the agent remembers.
Time-Travel Recall (recall_as_of)
Memory isn’t static. Agents need to reason about past states of their knowledge.
v0.5.0 adds recall_as_of (exposed via HTTP, gRPC, and both JavaScript & Python SDKs). Because Mnemara maintains versioned records in its backends (file and sled), you can now query the memory graph as it existed at any previous point in time.
Use cases include:
- Debugging why an agent made a particular decision days ago
- Compliance and audit trails
- Simulating “what would the agent have known at time T?”
- Building temporal reasoning capabilities on top of episodic memory
This is a significant step towards agents that don’t just have memory - they can navigate it through time.
Append-Only Changefeeds
Observability and reactivity just levelled up.
v0.5.0 introduces append-only changefeeds. You can now subscribe to a stream of memory events (writes, updates, deletions, compaction) through the admin API, HTTP, gRPC, and the SDKs.
This opens the door to:
- Real-time dashboards and monitoring
- Event-driven architectures where other services react to memory changes
- Building higher-level systems on top of Mnemara (e.g., memory-triggered workflows or audit logs)
Combined with the existing planner traces and provenance data, changefeeds make Mnemara one of the most observable memory engines available for AI agents.
Why These Features Matter for Trust
In Part 1 I argued that trustworthy agent memory requires more than just storage; it requires explainability, auditability, and evolvability.
v0.5.0 directly advances all three:
- The evaluation harness gives you a way to verify quality
- Time-travel recall gives you temporal auditability
- Changefeeds give you live visibility into how memory is changing
Together they move Mnemara closer to being infrastructure you can confidently run in production agent systems.
What’s Next
The roadmap has been updated to prioritise further work on the evaluation harness, richer watch/changefeed semantics, and continued improvements to time-based querying.
I’m particularly interested in feedback on:
- How you’d like to use the evaluation harness in practice
- Desired ergonomics for time-travel queries
- What kind of changefeed consumers would be most valuable
Try It Today
Mnemara v0.5.0 is available now:
- GitHub: deliberium/mnemara
- Crates.io and full documentation linked from the repo
You can use it embedded in Rust or run it as a standalone gRPC/HTTP daemon.
If you’re building agents that need memory they can truly trust, I’d love to hear what you’re working on and how Mnemara fits (or doesn’t yet).
Let’s keep building reliable long-term memory for AI agents - together.