Technical article

Project memory for AI agents: what should survive the chat?

Agents already use context. Real project work needs a memory layer that decides what matters, who can receive it, how current it is and how it can be verified before an agent acts.

Updated 2026-07-08

AI agents are no longer blank tools that wake up with no context. A serious coding agent can inspect files, read issue history, use tools, follow repository instructions and carry some local conversation state. The useful question is no longer whether an agent can use context. It can.

The harder question is which project memory an agent should actually use, who is allowed to receive it, how current it is, and how the team can verify what the agent saw before it acted.

The memory questions

More context is not the same as better project memory.

What should survive this chat?
What is relevant to the current agent task?
What is old but still authoritative?
What is recent but only tentative?
What came from a stronger source?
What contradicts or weakens a previous decision?
What can be shared with this audience?
What must be hidden?
What can be verified later?

Context is not project memory

Context is assembled for one interaction. Project memory has to survive interactions, preserve source and age, and remain usable when a different agent or teammate returns later.

Time is part of truth

A recent note often matters, but it should not automatically override an older requirement, security constraint or architecture decision with stronger authority.

Authority and relevance differ

A security rule may be authoritative but only relevant to certain changes. A debugging note may be relevant today but should not become project truth without confirmation.

Sharing is a memory problem

Partners, support teams and agents need different views of the same project. A useful memory layer should make safe scoped sharing easier than raw context dumping.

Example

One project, several safe memory views.

A backend agent, an external partner and a support lead can work from the same underlying project memory without receiving the same memory dump.

Backend coding agent

Receives
current endpoint behavior, source references, failure modes, required tests and active security constraints
Does not receive
support wording, partner negotiation notes and unrelated customer material

External partner

Receives
endpoint behavior, retry semantics, idempotency expectations, migration notes, expiry and provenance
Does not receive
private pricing, internal roadmap tradeoffs, raw security notes and customer-sensitive data

Support lead

Receives
customer impact, approved wording, rollout status and known limitations
Does not receive
raw backend debate, internal risk notes and implementation-only traces

SPM model

Durable, shareable, governed and internally smart memory.

SPM preserves requirements, decisions, completed work and source evidence across agent tools, then evaluates what matters, what still holds and what each authorized audience may receive through MCP, API or CLI.

  • durable project memory with source-linked summaries, decisions, policies and completed work
  • LLM-assisted memory triage that structures, summarizes, tags, relates and prioritizes mixed input
  • temporal memory across original intent, working state, current truth and history
  • authority-aware prioritization by source, role, confidence, relevance and temporal validity
  • queryable context graphs across requirements, decisions, sources, policies, actions and shares
  • scoped context packs for agents, teams, partners and support flows
  • governed sharing with permissions, expiry, revocation, entitlement logs and audience boundaries
  • provenance, hashes and pack verification before agent use
  • agent hardening with preflight checks, required tests, approvals and post-action reports

What to try first

Start with one project and one concrete handoff.

A coding agent receives a pack before changing a sensitive area.
A partner receives a scoped integration pack with exclusions.
A support lead receives approved customer language from engineering memory.
A later agent asks what changed and receives a temporal answer, not a flat transcript summary.
A preflight check requires tests or approval before a risky action.

Closing thought

Agents will keep getting better at reading context. That does not remove the need for project memory. It raises the bar for it. As agents do more consequential work, the memory they use should know what matters, what is still valid, what came from a stronger source, what can be shared, what must be hidden and what can be verified.