Guides/Memory
Persistent Memory for AI Agents: Typed, Linked Knowledge Your Agent Can Search
ContxOS gives agents long-term memory made of typed, linked entries (decisions, results, plans, bugs) that they search by meaning and drill into for the full story.
Jose R. Gonzalez, CPA · Founder of ContxOS
Updated · 2 min read
In short
In ContxOS, memory is a knowledge base of typed entries: decisions, results, plans, TODOs, bugs, findings, notes and more. Each entry is linked to the entries it supersedes, resolves, implements or supports. Agents search it by meaning, see how fresh and how trustworthy each result is, and open any entry to read its full body and the chain around it. The memory belongs to your project, not to any one chat or client.
Why chat memory is not enough
Assistant memory features remember facts about you, such as your name, your preferences and your job. That helps a conversation. It does not help a project. A project needs the decision you made last Tuesday, why you made it, what it replaced and what is still open, and it needs every agent working on the project to see the same record.
ContxOS stores that kind of knowledge, in a form an agent can query instead of re-reading.
What gets remembered
- Decisions, with their rationale.
- Results: what was completed and how.
- Plans and TODOs, with priority.
- Bugs, blockers and findings.
- Notes and documents worth keeping.
- Workflows: step-by-step procedures the agent can follow (see the workflows guide).
- Consolidated patterns and principles distilled from all of the above over time (see the consolidation guide).
Entries carry an authority level, working or authoritative, so an agent can tell a settled rule from a draft.
How agents find it
- Search by meaning. A question like "how did we handle the refund edge case" finds the relevant decisions even if they never use those words.
- Read the signals. Every result shows its creation date. Results that a newer entry has replaced, that have expired, or that contradict another entry are flagged and ranked accordingly, so the agent does not act on stale knowledge.
- Unpack. Opening a result shows its full body, its parent (for example the session or handoff it came from), its children and its links.
- Short phrases. Common tasks start from a two-word phrase ("resume", "new handoff", "next steps", "dream status") that routes the agent straight to the right procedure.
How it gets written
Your agent writes to ContxOS when you close a session ("finish session" or "new handoff"), when you tell it to record something, or as part of a workflow step. Nothing is captured silently. What is in the knowledge base is what you and your agent chose to keep, which is why it stays useful.
What to ask your agent
With ContxOS connected to Claude or another MCP client, requests like these work as written:
Search ContxOS for every decision about authentication and tell me which ones are still current.
Record this as a decision: we store amounts in cents, never floats, because of the rounding bug in the export.
What did we learn the last time the nightly sync failed?
Open that result and show me what it replaced.
Limits and requirements
- ContxOS remembers what agents write to it. It does not import chat transcripts or watch your files.
- Search quality depends on what was recorded. Short, specific entries with clear titles retrieve best.
- The free plan has a monthly write allowance; paid plans raise it (see Pricing).
Frequently asked questions
Is this the same as retrieval-augmented generation (RAG)?
It overlaps. RAG usually indexes documents you already have. ContxOS is a memory the agent writes as it works, with types, links and freshness signals that plain document search does not have.
Can I see and edit what my agent remembered?
Yes. Search and browse it in the ContxOS dashboard, or ask the agent to edit, relink, merge or retire any entry. A retirement requires a reason, which is kept.
Does the memory work across different AI models?
Yes. The memory lives in ContxOS, not in the model, so any connected agent reads the same knowledge.
Give your agent a memory that lasts
Connect ContxOS to Claude in about a minute. The free plan includes one project.