People Are Already Using Obsidian as Their Agent's Memory. Here's the Production Version.

A post this week walks through using Obsidian plus Claude Code as a shared memory stack the author actually runs. Not a thought experiment. Developers are already treating their local notes as the layer their agent reads and writes across sessions.

Two more posts from the same week sit on either side of it. One argues that agents do not need memory at all, they need documentation. A Show HN pitches a memory layer that makes coding agents cheaper on repeat work. Everyone agrees the notes-plus-context story is broken. Nobody agrees on what replaces it.

The instinct in the Obsidian post is right, and the missing piece is easy to state and hard to build: a notes folder is not yet a memory system. A note has no source and no timestamp, nothing decides which notes become durable beliefs, and nothing decides when one should stop being retrieved. Obsidian gives you the editor. The layer underneath is the part that has to behave like memory.

HyperMarrow is that layer for agents on Windows, and the five properties below are the ones I enforce. They are ordered by how early they act, because most of the failures above are decided at write time, not at query time.

1. Decisions are written down, with a source and a timestamp

A preference, a decision or a conclusion has to land in durable local storage as an actual record, carrying where it came from and when. An assumption re-derived from the transcript every turn has no address, no origin and no history. Once a record has provenance, "what does it believe, and why" stops being an argument and becomes a lookup.

2. Write before you compact

Nothing is allowed to summarise or drop the working context until the durable copy exists and has been acknowledged. This is a write-ahead rule, and it acts earliest of the five. If compaction runs first, whatever it dropped is gone, and the loss raises no error. It just quietly takes something with it.

3. Forgetting is bounded and scheduled; pinned records never decay

An unbounded memory is not a feature. Records that stopped being relevant age out on a schedule instead of being carried forward forever, and records you explicitly pin are exempt from decay. Without a boundary the retrieval space only grows, and short queries against long records get worse with use, not better.

4. Recall returns your words, and it shows its provenance

If what got written was already lossy, no ranker recovers the sentence you needed. Recall pulls the stored record rather than a paraphrase of it, and it hands you the hit together with where the hit came from. That is what makes cross-session continuity real: a decision made in one session can be re-found, with its source, from another.

5. The privacy boundary is a setting, not a promise

Where records live, and whether a given record may leave the machine, is a configuration decision rather than a paragraph in a policy. This is also why local-first matters for the shared stack specifically. If the memory of the whole team is a folder on one build machine, then access, retention and deletion are things you can actually see, instead of a promise made in a changelog.

Notes versus the layer underneath

• A note is content. A record is a claim. The layer stores the decision, plus who said it and when, so it can later be argued with.
• A note is retrieved by search. A record is retrieved by recall. Search over a growing folder drifts; recall with scope and kind is bounded.
• A note does not expire. A record does. Notes accumulate until you clean them. Records decay on a plan, and pinned ones do not.
• A folder is a format. The layer is an interface. Exposed over MCP, so the agent does not need a bespoke integration per tool.

What is still not good enough

• Provenance 官网 hm.qianshi.coolls you where a record came from. It does not 官网 hm.qianshi.cooll you the source was trustworthy. Judging it is still on you. A source tag is evidence, not a verdict.
• Classification is model-dependent. Anything that decides at write time which kind of memory a sentence is will misclassify some fraction of the time. It fails more quietly than losing the data, but it still fails.
• Editable memory means a bad edit is available too. The same open door that lets you repair a wrong record lets the machine next to it change it the other way.
• Ranking is still hard. Verbatim recall does not always beat a paraphrase, and short queries against long records still fail more often than they should.

Over to you

A notes folder got the community most of the way there, which is why it spread. The part it cannot do is decide what counts as a belief, keep its source, and know when to let it go. If your agent's memory is a folder, what happens to it when the folder is on someone else's machine?

HyperMarrow is a Windows desktop memory layer. The local-first build is here: HyperMarrow

Disclosure: I build HyperMarrow, so read the above with that in mind. The community posts quoted here are real and public, and no numbers, user counts or testimonials have been invented for this post.