Extended memory systems compared

Hermes plugs into nine extended memory providers, plus the plain-markdown control arm. What each one costs, where each keeps your facts, and which one fits which situation.

My agent’s memory started failing in the worst way: not forgetting, but unable to find what it knew. The agent that helps me run this site keeps months of accumulated facts about the publication in Mnemosyne, an extended-memory provider. The store sat nowhere near capacity. Recall was what degraded: as the facts piled up, the search got worse.

That wall is what the extended-memory market sells ladders against. Hermes can hand its memory to nine external systems. Nobody has put the field side by side; this piece does: the whole menu plus the do-nothing control arm, on five axes, ending in a pick. Picture each option as an archive building: the construction job is the install weight, the ground is where your facts live, the archivist is recall, the rent is the cost, the dusting is maintenance. The field differs on all five; star counts make bad maps.

The control arm: do nothing

Most readers never hit that wall, so the first option is refusing the market. The built-in store is free, local, already running, and stays on whatever else you plug in. hermes-09 fenced off what belongs in it: facts, not procedures. If your agent’s whole world fits on a few pages and you are its only user, the notebook wins, and what follows is a map of exits you have not needed yet. The wall to watch for is recall: the fact is in the store, and the search cannot find it.

The map: what can actually be plugged in

For everyone else, nine external providers can hold the memory instead. The provider comparison counts nine: eight bundled, plus Memori, which arrives via pip, Python’s installer. A tenth lives off the menu: Mnemosyne, the provider my own agent runs on. It is local-first, and the docs do not mention it.

Two commands tell you where you stand. hermes memory setup opens a picker of your install’s options. hermes memory status names the active provider and each auth mode: anything marked local needs no key. One external provider runs at a time, set through the picker or memory.provider in config.yaml, with the built-in store on beside it. Pick wrong, switch back, notebook intact.

The buildings on your own land

Ground first. Three bundled providers and one pip install keep every fact on your hardware and charge no rent. Holographic is the smallest surface in the set: bundled, two tools, a single-file SQLite database, zero dependencies (NumPy is optional). Every fact carries a trust grade, so a contradicted entry loses arguments to a confirmed one.

OpenViking, from Volcengine, is the opposite temperament: a self-hosted server, meaning the software runs on your machine, not theirs. Memory sits in a filesystem hierarchy and loads in tiers, so the agent reads a folder’s label before it opens the pages. Six tools, free, AGPL, and 35,000 stars eight months after launch: a number that moves that fast is what a ByteDance-sized backer does to a star count.

ByteRover works the moment before disaster. When a conversation grows long, Hermes trims it for room, and the Scribe copies the important pages into a portable knowledge tree first. Local-first, optional cloud sync, npm or a curl script, three tools.

And Mnemosyne, the off-menu tenth from the opening: pip-installed, layered recall separating working notes from long-term memory.

The storage companies

The cloud row is for facts allowed to leave the building. If your agent knows things other people trust it to know, this row is closed to you; for everyone else, the free tiers are real. Mem0 is the famous one: 64,500 stars as of September 2026, the only household name on the menu. You hand it boxes; a server-side model labels and retrieves them. Cloud, self-hosted Docker, or the open-source core in-process, via pip and an API key.

Supermemory offers cloud or a local branch you run yourself. 29,000 stars, MIT licensed, pitched on context fencing: a machine fact never wanders into a people file. Memori is cloud-only in Hermes and arrives with pip install hermes-memori. 16,000 stars. Its angle is tool-aware memory: which tools did what.

RetainDB is less a store than a finding service: vector search matching records by meaning rather than exact words, keyword matching, a rerank pass, ten tools, and delta compression, which sends only what changed. The youngest thing here: 54 stars and five months old, a commercial service rather than an open-source movement. Pricing when this was written: free, $9, $29 and $99 a month on its pricing page.

The cross-referencers

Local buildings and storage companies answer one question: find the fact. When the job is to connect the facts, two of the field think in links instead of pages. Hindsight keeps a knowledge graph, facts as linked cards you can walk between. It runs on their cloud or on PostgreSQL, a free database engine on your own machine. MIT licensed, 22,000 stars, three tools. Its reflect tool answers with connections: ask it to think across the whole corkboard rather than pin one card.

Honcho keeps a model of people instead of facts about them, maintained across sessions and agents. 7,000 stars, AGPL, cloud or self-hosted. Several agents dealing with the same humans can share the biography.

Where this row earns its keep: two of your agents each hold half of one fact. One knows the person; the other knows the thing the person looks after. No amount of lookup joins those halves. A link does.

The names everyone already knows are missing, and the biggest is claude-mem at 93,000 stars, the largest repository in the space, unable to plug into Hermes: it hooks into Claude Code’s session lifecycle, compressing and reinjecting context, which is not a pluggable provider API. The same goes for Cognee at 30,000 stars, Zep and Graphiti at 30,500, the older Zep Community Edition deprecated, Letta at 24,500, memU at 14,000, and MemOS at 11,000. The reason is shape: fame tracks platform fit, and none of these fits this platform’s provider socket.

How to choose

A tour is not a decision. The order that turns one into the other: boundary first, shape second, rent third, fame never.

Count what you have: hermes memory status. Built-in active and searches still working? Stop and enjoy the money.
Draw the data boundary: mark anything that should never stand on someone else’s land. One marked line removes every cloud option, which is why this step comes first.
Choose the recall shape: notebook pages, linked cards, a model of people, or a specialist finder. Most single-agent homes want the first.
Price the rent against the dusting. Cloud is a monthly line that grows with use; local is your electricity, your disk, and a backup habit you keep.
Switch one dial: hermes memory setup, pick one. A wrong choice costs you a switch back.
Re-audit after a month of real use. If recall did not improve where it hurt, the rent is not buying anything.

The verdict

Four situations, four answers:

If the facts fit a few pages: the notebook. It is the default because it is enough.
If any fact would embarrass you in a breach: build on your own land. Holographic for the smallest surface, OpenViking for structure if you can run a server, ByteRover if your worry is compression throwing things away, Mnemosyne for layered recall. Discount that last one by this site’s conflict of interest.
If the facts are boring and you refuse to run a server: the storage companies. Mem0 for the mainstream path, Supermemory for the self-host exit door, Memori for tool-heavy workflows, RetainDB if search quality is the entire job and a five-month-old vendor does not scare you.
If more than one agent shares one world: the cross-referencers. Honcho for a shared model of the people, Hindsight for shared linked facts, the local Postgres mode if your step-two boundary says so.

Honest limits

There are no benchmark numbers here; the omission is deliberate. The same product has been scored at 49% in one published run and 94 in another, depending on who counted, and the comparison blogs themselves warn that vendor-counted benchmarks are marketing. No system on the list has an independent benchmark anyone trusts, so the ranking rests on what you can check: architecture, storage, cost, your situation. Stars measure attention, not fit. Pricing moves; the tiers quoted were live when this was written.

Takeaway

Pick in the order above: what must never leave your control, then the shape of recall, then the rent. When you are ready to wire one in, the installs live in the Starter Kit, not this piece: a comparison should end at the decision, not the plumbing.