Đoạn xem trước bài giảng

AI Agent Memory Between Sessions: Session Handoff, PROGRESS.md and Bootstrap Contract

How an AI coding agent remembers a project across sessions: compaction vs context reset, session handoff, PROGRESS.md, git checkpoints, decision log, warm start and init.sh. Context recovery cost metric.

cross-session continuity session handoff bootstrap contract warm start

Bạn sẽ học gì trong bài giảng này

Implement continuity artefacts and reduce re-exploration time to almost zero.
Design a bootstrap contract that verifies the agent's readiness rather than relying on guesswork.
Measure recovery cost before and after — and see the difference in hard numbers.

Phần đầu bài giảng

Fifteen minutes to remember yesterday

An engineer arrives each morning with their memory completely wiped
They're brilliant, but can't recall yesterday's decisions
They have to re-read the code and piece together the history
An agent that has lost its context behaves exactly the same way

Picture an engineer who turns up at work every morning with their memory of the previous day completely wiped. They're brilliant, but they don't remember which architecture option was chosen yesterday or why the others were ruled out, which files have already been rewritten, or where the tests stand. They have to re-read the code, reconstruct the history, and only then carry on working. This is exactly how an agent behaves once it's lost its context: fifteen to twenty minutes go on reconnaissance, and decisions get made all over again — often against what was already settled yesterday.

The agent repeats its recon and argues with its past self

Thirty minutes of work — and the context runs dry
The new session knows none of the decisions, changed files or test status
Fifteen minutes go on repeating the reconnaissance
The cure isn't more memory — it's a logbook

Thirty minutes of work, and the agent is already arguing with itself. It had built most of the feature, then the context ran out, and the new session doesn't remember which decisions were taken, why option B was chosen, which files have been changed, or where the tests stand. It spends fifteen minutes re-running the reconnaissance and, more likely than not, goes against the previous approach. The right mental model here is a brilliant engineer with total amnesia every morning. And the fix isn't giving the agent more memory — it's introducing a logbook: a set of files that capture the state between sessions.

Why growing the context window doesn't solve the problem

Even a million tokens get exhausted by complex tasks
Code, decision history, tool output, conversation — it all grows
Needs grow faster than windows do
This is an architecture problem, not a capacity question

Think of an exam: no matter how many hours you're given, a genuinely hard question will still eat the time right up to the last minute. It's the same with an agent's context window: give it a million tokens if you like — a complex task will exhaust that too. The agent reads the codebase, accumulates a history of decisions, takes in tool output and holds the whole conversation, and all of it grows faster than model windows do. You can't buy reliability simply by raising the token limit — sooner or later the exam clock runs out, and the question isn't whether that moment arrives, but what you've prepared for when it does.

Bài giảng đầy đủ dành cho thuê bao

Còn 24 slide nữa trong phiên bản đầy đủ.

Video, toàn bộ slide, ghi chú, bài kiểm tra, gia sư AI và phỏng vấn sẽ mở sau khi đăng nhập và đăng ký thuê bao.

Đăng nhập hoặc đăng ký Đăng ký thuê bao

Nội dung bài giảng

Fifteen minutes to remember yesterday
The agent repeats its recon and argues with its past self
Why growing the context window doesn't solve the problem
Compaction keeps the 'what' but loses the 'why'
Context anxiety: the agent rushes like a student racing the bell
❓ Checkpoint: compact or reset?
What compaction keeps — and what a reset wipes clean
A practical criterion: the sixty percent mark of the window
💡 Walkthrough: forty percent and a single session — that's compaction
The four continuity artefacts
PROGRESS.md — where we are now
DECISIONS.md — why we did it this way
Git commits as free checkpoints
Fifteen minutes versus three: work out your cost of amnesia
Fifteen minutes versus three: work out your cost of amnesia
Four fields that are enough for a complete handoff
Coming up next: initialisation as a separate phase
🛠 'Add search' — and the agent has already spent half an hour getting to grips with the project
Implementation and initialisation optimise different things
Pitfalls: code accumulated without verification
The bootstrap contract: four readiness conditions
Myth: initialisation is a waste of time at the start
Myth: initialisation is a waste of time at the start
Approach A versus approach B: who wins over four sessions
Approach A versus approach B: who wins over four sessions
End-to-end project: step three — the payment backend logbook
From amnesia to discipline — and on to scope