Platform · Memory

Structured memory

Agents that run for weeks cannot start every session from a blank context. Structured memory gives a campaign a working set, a full record of what it has tried, the rules it has learned, and the tools it has built, so it resumes exactly where it left off.

01The problem

Take an autonomous synthesis lab like Berkeley Lab's A-Lab, which ran unattended around the clock for weeks, planning and running experiments day after day. A campaign like that only pays off if it remembers every prior result, every surrogate-model update, and every region already ruled out. Without persistent, structured memory it re-explores dead ends and throws away skills it already earned.

Memory tiers of an autonomous campaign. Take an autonomous synthesis lab running for weeks as an example. A layered memory keeps its active task, its full run history, the rules it has learned, and the skills it has built, all linked in one temporal graph.

02How we solve it

Four memory tiers, one temporal graph.

Agentriq gives every agent a layered memory: what it is doing now, what it has done, what it has learned, and what it can reuse, all indexed in a temporal knowledge graph.

Paging memory into a limited context. An OS-style controller keeps only what the current step needs in the context window and pages the rest to external stores, so nothing is lost as a campaign grows.

Four-tier cognitive memory (CoALA)

Working memory holds the active task while long-term memory splits into episodic runs, semantic facts, and procedural skills, following the CoALA taxonomy for language agents.

Agentic RAG retrieval and paging

Agentic RAG runs multi-hop retrieval over the memory graph while an OS-style controller pages results between the limited context window and external stores, so the agent pulls back exactly the memories it needs and nothing is lost.

Temporal knowledge-graph memory (Zep)

Every fact, event, and relationship is stored in a time-aware graph, so the agent retrieves not just what was true but when it was true and what has since changed.

Consolidation and novelty gating

Recurring outcomes are distilled into durable rules and only genuinely new information is written back, keeping memory compact and retrieval sharp over months.

03What it produces

Context that survives a months-long run.

Temporal-graph memory: more accurate, far faster. On long-memory benchmarks, temporal knowledge-graph memory such as Zep raises answer accuracy while sharply cutting retrieval latency versus a full-context baseline. The figures shown are the published results, used here to illustrate the pattern.

An agent resumes a months-long campaign with full context, never re-exploring ruled-out regions and reusing its own tools; temporal-graph memory approaches such as Zep report up to 18.5% higher accuracy on long-memory benchmarks while cutting retrieval latency by roughly 90%.

04The agentic loop

Nothing learned is ever lost.

Agentriq wraps memory in a closed agent loop, so every new result is verified, explained, judged, and written back into the right tier for the next session to build on.

From work to a verified decision. Every new result is verified, explained against prior runs, judged for novelty, and written back to the right memory tier.

01Load the active task, its constraints, and the relevant history into working memory at the start of each session.
02Run the work, then a verification agent checks every new result for consistency and convergence before anything is trusted.
03A root-cause agent and Agentic RAG explain the result by correlating it with prior runs across the temporal knowledge graph.
04A judge agent decides whether the finding is novel and durable enough to keep, or a dead end to gate out.
05Write the outcome back to the right tier, an episodic run, a semantic rule, or a procedural skill, so the next session starts ahead.

05Tooling

Simulation, solvers & frameworks.

Get started

From engineering problems to autonomous solutions.

Talk to our engineers john.ferderigos@agentriq-sa.com