Session Log vs Memory vs Trace
How USL differs from Mem0-style memory and observability telemetry
A session log, an agent memory system, and an observability trace may all ingest conversation data, but they preserve it for different reasons.
| layer | optimized for | typical output | source of truth? |
|---|---|---|---|
| USL session log | exact history, provenance, replay, migration, and audit | ordered canonical events plus native evidence | yes, for captured session history |
| memory layer (for example Mem0) | retrieving useful facts and preferences for a future prompt | extracted, updated, deduplicated, or embedded memories | no; it is a derived knowledge layer |
| observability trace | diagnosing latency, errors, cost, and service behavior | spans, metrics, logs, and attributes | no; it is operational telemetry |
Why Mem0 having a session_id is not the same thing
Mem0 can scope memories by user, agent, application, run, or session and can retain raw messages when configured. Its primary abstraction is still a memory: information selected or stored for later retrieval. USL's primary abstraction is an append-only record with sequence, integrity framing, canonical event meaning, and evidence back to the native source.
That difference matters when you ask:
- Can I prove the order and origin of tool calls and results?
- Can I preserve an unknown vendor payload without understanding it?
- Can I rebuild projections after a crash or schema change?
- Can an exporter determine whether a native harness can safely resume?
Those are session-log questions. “What preferences or facts should the next prompt recall?” is a memory question.
Use them together
The recommended architecture is additive:
native harness / ACP stream
|
v
ASP canonical events + evidence -> USL durable append log
| |
| +-> telemetry traces and metrics
+-> redacted chunks -> memory extraction / semantic retrievalThe USL log remains the rebuildable history. A memory index can be regenerated under a new extraction or redaction policy, and telemetry can be retained under its own operational policy. A memory hit or trace span should carry evidence or correlation IDs back to the session rather than replace it.
Current boundary
USL currently ships durable append/read/verify primitives and harness conversion components. The richer full-text, vector, lineage, analytics, and subscription system in the query RFC is a target architecture, not a claim that all features ship today. See Query Plane and Search & Indexing.
References: Mem0 repository, OpenTelemetry generative AI conventions, and the ASP specification.