When to add a state integrity layer (and when observability is enough)
If you run agents across sessions and tools-IDE, chat, cron, CLI-you already pay the tax of re-prompting, manual sync, and homegrown markdown or JSON workarounds. Logging what ran is a solved problem. Knowing the composed, replayable state of each entity (contacts, tasks, money facts, decisions) across those writers is a different problem. This page separates the two.
Where are you?
How agent memory infrastructure evolves
Most teams follow the same progression. The pain that drives adoption of a write-integrity layer is invisible at first and unavoidable later.
Memory bolted onto Postgres, Redis, or a vector store. Works for simple use cases. Nobody can answer what the agent learned, when, from what source, or whether it contradicted last week’s state—but the system doesn’t visibly break.
Dedicated memory abstraction focused on retrieval quality: right context, right time. Solves legible pain (bad recall, bloated context). Leaves the write path unaudited—LLM-mediated summaries treated as ground truth with no provenance.
Agents move from low-stakes assistants to high-stakes actors: money, procurement, compliance, long-running autonomy. The question shifts from “did the agent retrieve the right thing?” to “can I prove what the agent knew, when, and whether that knowledge was legitimate?”
Teams that adopted append-only, schema-constrained state early have a compounding advantage: every agent write is traceable from day one. Teams that retrofit have a gap in their audit history—everything before the migration is a black box.
Building multi-agent systems? See how shared state accelerates the inflection point.
Already solved
What existing tools cover well
Audit and observability have mature solutions. If your requirements stop at knowing what happened (who did what, when, and in what order) you do not need Neotoma. These are real, valuable capabilities, and they are handled.
Append-only event streams with structured payloads. Datadog, Splunk, ELK, or a Postgres table: recording what happened is a solved problem.
Real-time visibility into agent behavior: metrics, traces, anomaly detection. Mature tooling covers this across every stack.
Show users or operators what's happening in their systems. Chronological, filterable, exportable. Standard product feature with standard solutions.
On-demand or scheduled data exports for auditors. Query by entity and time range, package as CSV or PDF. Straightforward with any data store.
The gap
Self-assessment
By entity type
The real question
Observability tells you what happened. State integrity proves what was true.
What observability gives you
- A complete record of every event, decision, and action in your system
- Real-time dashboards and alerting on agent behavior
- Searchable logs by entity, agent, time range, and event type
- Exportable audit trails for compliance and incident response
If this is all you need, you're set. Existing tools and straightforward engineering cover these requirements well.
What state integrity adds
- Deterministic reconstruction of entity state at any historical point
- Multi-writer conflict resolution with guaranteed consistency
- Every decision bound to the exact policy version and context that produced it
- Provenance chains traversable from outcome back to origin
- Cross-system entity correlation into one canonical snapshot
This is the gap: not a harder version of observability, but a different requirement that existing tools do not address.
Integration model
Between your agents and your database (not instead of it)
Your existing Postgres, MySQL, or managed database stays as the system of record for business data. Neotoma adds a layer for a category of data that did not exist before agents started writing autonomously: observations, inferences, entity resolutions, and decisions—with append-only provenance, schema constraints, and deterministic reduction.
Every day without write integrity is a gap in your audit history
Neotoma is an open-source state integrity layer: deterministic temporal reconstruction, multi-writer consistency, and version-bound provenance. Install in five minutes. Everything your agents write from that point forward is traceable, auditable, and consistent. Everything before is a black box.