giving an ai coding agent memory of your actual work

ai coding agent memory is not a model problem. every agent you use starts each session from zero, and the fix is a record of your work rather than a bigger context window.

an agent in your repo can read every file, every commit, every test. it cannot read the thursday you spent proving that the retry logic was fine and the clock was wrong. that afternoon left no artifact. so on friday you explain it again, and the agent believes you, and neither of you can point at the evidence.

here is what the gap actually consists of, as of september 2026, and what closes it.

why does an agent start every session blank?

because the session is the memory, and the session ends. call it session amnesia and it stops sounding like a defect.

a language model holds no state between calls. everything it knows about your work arrives inside the context window at the top of a turn: your prompt, the files its tools read, the output of the commands it ran. when the process exits, all of it goes. the next run rebuilds from whatever is on disk.

that is not a bug anyone forgot to fix. it is the architecture, and it is why coding agents forget in the same way across claude code, codex and cursor. what differs is how much scaffolding each one puts around the blankness. when people say claude code memory they usually mean that scaffolding: a rules file, a session summary, a directory of notes. useful, and not the same thing as a record of what you did.

the industry response has been context engineering, which is the practice of deciding what goes into the window and in what order. the term became standard in practitioner guides once prompting alone stopped solving production problems. it is real work and it is worth doing. it also cannot conjure a fact that was never recorded.

what the agent cannot see

four categories, and they are all the same category underneath.

commits are a lossy format on purpose. they record what changed and discard why, because a diff is a statement about code rather than about people.

so the agent inherits a codebase with the reasoning stripped out. then it reasons from first principles, and produces something confident and slightly wrong.

is a claude.md file memory?

no, it is a briefing, and the difference matters more than it sounds.

a claude.md file, a cursor rules file, an agents.md, a project readme aimed at a model: all of these are hand written summaries of what you thought was important on the day you wrote them. they are worth having. teams that keep one get noticeably better sessions than teams that do not.

but a briefing has the same failure mode as every other manual note. it is written once, at the moment of highest enthusiasm, and updated never. by week six it describes a system that has moved. the agent then follows it confidently into a wall, which is worse than having no file at all, and it is the same decay we go through in why your notes app was never going to work.

agent onboarding fails in the way human onboarding fails. the document is accurate and the work has moved on.

what actually counts as memory for an agent

three properties, and a source has to have all three.

1. it records without being asked. anything that needs a person to file it will stop being filed. this is the whole reason meeting transcription stuck where team wikis did not.
2. it is queryable, not just stored. a folder of screenshots is storage. text you can search, filtered by day and by app, is a source.
3. it stays under your control. an archive of your work is more sensitive than the repo, so where it lives is part of the design rather than a footnote.

the first property is the one people underestimate. per the anthropic economic index for may 2026, 51.38% of sampled usage is augmentation, meaning long iterative sessions with a person in the loop rather than one shot answers. the dominant mode is a person and a model working a problem over hours. the model forgets the hours.

the same dataset says something sharper when you look at what people ask for.

software development is 11.51% and searching your own organisation's knowledge is 3.61%. the top individual work task across all of it is searching electronic sources for information, at 4.95%, with reference searching second at 3.74%. retrieval is what these tools mostly do. your own work is the one corpus they cannot retrieve from.

what changes when the agent can query your week

the failure mode flips from confident invention to a checkable answer.

concretely, these are the questions that go from unanswerable to answerable inside a session:

• "what was i doing on august 14, 2026, before this branch broke?" the apps, the docs, the pages, in order, rather than a reconstruction.
• "find the error text about the keychain i saw last week." searchable text from what was on screen, including text inside images and calls, which no file search reaches.
• "we already ruled out one of these three approaches, which one?" the thread where it was ruled out, instead of a fresh guess.
• "what did the vendor doc say before they rewrote it?" the version you actually read.

none of that requires a smarter model. it requires a source. a small model with good retrieval beats a large model guessing, and it costs a fraction as much, which is the argument we made in how to give an ai agent context about your own work.

how do you connect a record of your work to an agent?

two shapes exist in 2026, and they are worth telling apart.

the first is a tool interface. the model context protocol, published by anthropic in late 2024 and adopted broadly by 2026, lets a host application call an external source during a session. that is the general standard, explained in what is an mcp server.

the second is simpler and needs no protocol at all: an app that hands your question, plus the relevant slices of your own history, to the coding agent you already have installed on the machine. the agent runs where it always ran. the app supplies the context it was missing.

this got easier through 2026 because the agents moved onto the desktop. anthropic gave claude mouse and keyboard control of the mac on march 23, 2026, openai shipped desktop computer use in codex on april 16, 2026, and apple announced siri ai with onscreen awareness and personal context on june 8, 2026. microsoft went at the same gap from the windows side with recall. we walked through the desktop agents in computer use agents on a mac.

the last two are not rivals. they are the same idea reached from different directions, which is that the agent should be able to look something up rather than be told everything up front.

what to check before you point an agent at your history

three questions, in this order, and they apply to any source you connect.

where does it run? a source that answers from a database on your own disk is a different risk profile from a hosted connector holding tokens for your accounts. ask before you install.

can it write? read only is the safe default for an archive of your own past. a source that can also modify things has a much larger blast radius when a model is wrong or when instructions arrive from a page it read.

what can reach it? anything an agent reads can carry instructions aimed at the agent. the cloud security alliance's 2026 research note documents indirect prompt injection in the wild, and academic testing has found browser agents vulnerable 24 to 47% of the time. the background is in what is prompt injection, and the general risk list is the owasp top 10 for llm applications.

exclusions are the other half of that answer. an archive that cannot exclude a password manager, a banking site or an hr portal is not one you should be running on a work machine.

where remynd fits, and where it does not

it is the record, not the agent, and the boundary is worth stating exactly.

remynd captures the focused window on a mac, runs ocr locally so what you looked at becomes searchable text, and keeps the index on your machine. recordings default to 30 days of retention, you can exclude specific apps or sites from capture entirely, and your history is read only everywhere in the app.

call transcription runs on device using an mlx speech model on apple silicon. it is a setting you turn on per call rather than something running by default, and announcing the recording is on you rather than on the software.

there is an open agent window that drives the claude or codex cli you already have installed on your mac, handing it your own history as context. it is not a hosted service and it does not install a model. if you do not already run one of those tools, that window has nothing to drive.

what leaves the machine, plainly: sign in, crash reports, and any question you ask a cloud model, which sends the retrieved slices with it. capture and storage stay on your own mac or your own private encrypted bucket. the fuller version is in private ai on your mac and on for engineers.

the honest summary

the bottleneck moved, and most teams are still optimising the old one.

for two years the answer to a bad session was a better prompt, then a bigger window, then a better model. all three helped. none of them addressed the case where the fact you needed was said out loud on a call, or read once in a browser tab, and written down nowhere.

fix the record and the model gets easier, not harder. that is the same conclusion we reached from the human side in why you forget what you read, and from the tooling side in what is an ai agent.

download remynd for mac and give tomorrow's session something to read.