AI project memory
AI memory remembers facts. A project needs more.
Preferences, durable facts, details worth recalling next time: AI memory is genuinely useful. But a project that runs for months accumulates something memory was never designed to hold: a current state, a history, and the reasons behind both.
What AI memory does today.
Most AI tools now remember things between conversations, and that solves a real annoyance: you stop repeating yourself. Typically, memory holds three kinds of things.
- Your preferences: tone, formats, how you like to work.
- Facts about you and your context that may be useful later.
- Recall inside one platform, so the same assistant feels consistent.
All of this is memory of the agent: what one assistant, on one platform, knows about you. It is worth having. It is not the memory of a project.
AI memory remembers facts. A project needs decisions, open work, documents and history.
Facts, state and intent are three different things.
A fact is stable: the product targets small teams. State changes every week: which tasks are open, what shipped, what is blocked. Intent explains why the state looks the way it does: what was tried, what was ruled out, and for what reason. Memory captures the first kind well. A long-running project runs on all three.
A project context should include the decisions and their rationale, the work still open, the documents that carry durable knowledge, and the history of how the project changed over time.
Project continuity is the ability to resume a project with its current state, history and rationale intact, regardless of which human, AI, tool or session picks it up next.
Four kinds of context, each with a job.
Decisions
What was chosen, why, and which alternatives were rejected. When a decision replaces an earlier one, the old one keeps its trail instead of silently disappearing.
Open work
Tasks and unresolved threads stay visible until someone actually closes them. The project knows what needs attention now, not just what happened before.
Documents
Markdown documents, versioned, with provenance: who changed them, from which tool, in which session. The knowledge stays part of the project itself.
Sessions
Each work session leaves a record: a summary, the decisions made, the documents touched, the tasks opened or closed. One state of the project connects to the next.
Decisions preserve the why. Open work preserves the present. Documents preserve the knowledge. Sessions preserve the movement.
Memory lives per platform. Projects do not.
What Claude remembers stays with Claude. What ChatGPT remembers stays with ChatGPT. Switch models for one task, hand the work to a teammate, or come back three months later from a different client, and per-platform recall stops helping. We look at how this plays out with Claude’s project memory and with ChatGPT’s project context in detail.
The context of a project should not care which tool opens it. That is the difference in one sentence:
AI memory helps an agent remember facts. Project continuity helps a project preserve its state, history and rationale so that someone or something else can pick it back up later.
How Palimpse holds a project’s memory.
Palimpse gives every session two moments of discipline, and asks almost nothing in between. At the start, your agent asks Palimpse to catch up. It receives a briefing: open tasks first, then what changed since its last visit, the durable facts kept in project memory, and the documents worth reading.
Open tasks
Verify the onboarding flow end to end. One search issue is still open.
Since your last visit
The pricing copy was rewritten and the positioning document changed.
Latest decision
The earlier pricing framing was ruled out, and the reason is on record.
Project memory
The landing voice is direct and concise. All UI copy is English.
At the end, the session is logged: a summary, the decisions with the alternatives that were rejected, the documents touched, and the tasks opened or closed. The next session, in whichever tool, starts from that state.
Palimpse is a standard MCP server. Paste its URL into claude.ai, Claude Code, ChatGPT, Cursor or any compatible MCP client, authorize in the browser, and the project’s memory follows the project rather than the platform.
Give your project a memory of its own.
Connect Palimpse once over MCP. Every session starts briefed and ends recorded, whatever model, conversation or teammate comes next. Free for solo.