Skip to content

Memory layer for agentic systems

Memory infrastructure for agents that need the right context

Entor gives AI agents a deterministic memory and retrieval layer designed for workflows where retrieving something similar is not good enough.

Retrieval
Structured traversal
Behaviour
Deterministic
Engine
Open source
traversaldepth 3 · 8 nodes evaluated
querypolicy.v4chunkamendmentchunkchunkclause 12bruling 2024

input

agent query

strategy

learned traversal

output

ordered context

The problem

Similarity is a ranking signal, not a correctness guarantee

Vector search is good at finding text that reads like your query. In workflows where an agent acts on what it retrieves, the harder requirement is returning the specific facts and relationships that apply to the case in front of it.

Conventional retrieval

  • Finds semantically similar content
  • Often treats memory as independent chunks
  • Can lose relationships and historical context
  • May return plausible but operationally incorrect information

Entor

  • Preserves structured relationships between memories
  • Traverses relevant context rather than performing only nearest-neighbour lookup
  • Supports deterministic and inspectable retrieval behaviour
  • Designed for agents making consequential decisions

How it works

A retrieval path you can follow end to end

Memory is written once with its relationships intact, traversed against the task at hand, and returned as an ordered set the agent can act on.

  1. 1

    Store structured memory

    Facts, events, relationships, and operational context are written into the memory layer.

    write(entity, relation, context)

  2. 2

    Traverse relevant context

    The engine evaluates connected memories and follows the paths relevant to the current task.

    traverse(task, depth)

  3. 3

    Return decision-ready context

    The agent receives a compact, ordered set of memories suitable for its next decision or action.

    context[] → agent

Use cases

Built for workflows where the wrong context has consequences

Entor is aimed at agentic systems in high-risk and context-sensitive domains, where precise historical and relational context determines whether an action is correct.

  • versioned obligations

    Compliance agents

    Obligations change over time and rarely live in one document. An agent reviewing a control needs the clause that currently applies, the amendment that replaced the old one, and the prior determinations attached to that entity — not a paragraph that merely reads like the rule.

  • dependencies and history

    Incident and infrastructure agents

    During an incident, the useful memory is relational: what changed, what depends on the failing service, and how a similar alert was resolved before. Chunked logs lose those links exactly when an agent is about to take a remediating action.

  • policy in force

    Insurance and claims workflows

    A claim decision depends on the policy version in force on the date of loss, its endorsements, exclusions, and the claimant’s earlier claims. Retrieving the wrong policy revision produces an answer that looks defensible and is not.

  • inspectable trail

    Legal and audit systems

    Legal and audit work is built on chains of reference and evidence. Reviewers need to see which memories were consulted and in what order, so a retrieval step can be inspected and re-run rather than trusted on faith.

Open source

Built on an open-source retrieval engine

Entor is built on Hippocampus, an experimental memory engine exploring learned traversal, structured context, and deterministic retrieval for AI agents.

The engine is public and under active development. If you want to understand how traversal is modelled before you depend on it, read the source.

retrieve.py

illustrative

# illustrative shape of a retrieval callresult = memory.retrieve(    task="assess claim 8841",    anchors=["policy:8841", "loss_date:2025-02-11"],    depth=3,) for memory_item in result.context:    print(memory_item.id, memory_item.path)

Interface shown for illustration. The managed API is in development and will be shared with early access teams.

Early access

Join the waitlist

Entor is in early development. We are working with a small number of teams running agents in context-sensitive workflows to shape the managed memory layer.

  • Tell us about your workflow and retrieval requirements
  • Early access to the managed memory API as it stabilises
  • Direct contact with the engineering team

We use these details only to contact you about early access.

Contact

Sales and design partnerships

If your team is building agentic systems in high-risk or context-sensitive workflows, we would like to hear how you handle retrieval today.