Trustworthy Agent Memory and Context
AI agents do not only need bigger context windows or more remembered text.
They need context that can be trusted.
This guide is about the operating layer around an agent: GitHub, Slack, Linear, Jira, Sentry, docs, incidents, deployments, decisions, permissions, provenance, traces, and evals.
The problem is simple to state and hard to solve:
How does an agent know what is true, current, relevant, allowed, and safe to act on?
Start here
Section titled “Start here”- Memory is not context
- Which context is allowed to win?
- Accountable context
- Provenance before persistence
- Stale context is a production risk
- Permissions are part of memory
- Evals for agent context
Concepts
Section titled “Concepts”- Inference
- Eval harness
- Embedding
- Retrieval and reranking
- Context window
- Tool calling
- Agent trace
- GraphRAG
- Vector search vs full-text search vs graph traversal
Design
Section titled “Design”Experiments
Section titled “Experiments”Research
Section titled “Research”- Provider context map
- External memory systems
- OptMem research notes
- Buzz research notes
- CocoIndex research notes
- HelixDB initial notes
Principles
Section titled “Principles”- Memory without provenance is a liability.
- Stale context is worse than missing context.
- Permissions are part of the memory model, not an integration detail.
- The system should make agents informed before it makes them autonomous.
- Durable memory should be easy for humans to inspect and correct.
- Claims about better agent behavior need evals, not vibes.
The current focus is software teams using AI agents and coding assistants.
That includes:
- source control, PRs, reviews, commits, and CI
- tickets, priorities, ownership, and acceptance criteria
- chat decisions and informal team knowledge
- docs, runbooks, ADRs, and onboarding notes
- monitoring, Sentry issues, incidents, and deploy history
- agent instructions such as AGENTS.md, CLAUDE.md, Cursor rules, and Codex skills
Applied work
Section titled “Applied work”These principles are being applied in Memoato, an open-source product that preserves raw evidence, separates reviewed facts from inferences, and exposes scoped recall through MCP.
Authorship
Section titled “Authorship”Written by Hrvoje Pavlinovic. Maintained through HILLS Lab.