I’m running openclaw v15 on Linux. I wish to improve its memory to make my bot a true companion with a capable memory for social interaction, but I can’t decide which memory systems to choose. I think honcho is key, so I have installed that. But how much will it benefit from additional systems like qmd and memory-wiki? And is openclaw-dreaming advisable since honcho already has a dreaming function?
#Which memory systems to choose
1 messages · Page 1 of 1 (latest)
Let me check the local docs and memory-related references so I don’t guess.
I found the current OpenClaw memory docs. I’m also checking Honcho’s side so I don’t hand-wave the overlap question.
Short version, if your goal is social companion memory, Honcho is the biggest win.
- Honcho: best for cross-session social memory, user modeling, and “this bot actually knows me”.
- QMD: great add-on for searching local notes, docs, and transcripts. Big for knowledge recall, smaller for pure social memory.
- memory-wiki: useful if you want a curated, inspectable knowledge base with provenance and dashboards. Nice, but not my first pick for companionship.
- OpenClaw dreaming: in current docs, this is the
memory-coredreaming system that promotes local Markdown memory intoMEMORY.md/DREAMS.md. Since Honcho already does its own extraction, summarization, and dreaming, I’d skip this at first unless you specifically want local, human-reviewable file memory.
My recommendation:
- Start with Honcho only
- Add QMD if you want better recall from files/docs/transcripts
- Add memory-wiki later if you want a maintained knowledge vault
- Add dreaming last, only if you want local Markdown consolidation
So for your use case, I’d rank them:
Honcho >> QMD > memory-wiki,
and dreaming is optional / probably redundant at first
One caveat, v15 is pretty old, so current memory docs may not map 1:1 to your setup. If you want exact version-safe advice, paste:
openclaw --version
Docs:
https://docs.openclaw.ai/concepts/memory
https://docs.openclaw.ai/concepts/memory-honcho
https://docs.openclaw.ai/concepts/memory-qmd
https://docs.openclaw.ai/plugins/memory-wiki
https://docs.openclaw.ai/concepts/dreaming
If you want, I can give you a best social companion stack and a best fully local stack next.
Yes please. BTW I’m running 2026.4.15, so the latest version.i have a local honcho database. Capture is done by a small local llm, subsequent processing by cloud models.
Nice, that helps a lot. On 2026.4.15 with local Honcho plus local capture / cloud processing, I’d go with this:
Best stack for social companion memory
1. Honcho as the core
- keep this as the main social memory layer
- it already gives you cross-session memory, user modeling, and semantic recall
2. Add QMD only for non-social recall
- docs
- notes
- worldbuilding
- project files
- maybe curated transcript archives
My suggestion: if you add QMD, use it mainly for files/knowledge, not as a second conversation-memory brain.
3. Skip memory-wiki at first
- add it later only if you want an inspectable knowledge vault
- great for stable facts, entities, contradictions, dashboards
- not the first thing that makes the bot feel more “social”
4. Leave OpenClaw dreaming off initially
- since Honcho already dreams, turning on OpenClaw dreaming too is likely more overlap than benefit at first
- I’d only enable it if you specifically want:
- local
MEMORY.md DREAMS.md- human-reviewable Markdown memory promotion
- local
My actual recommendation for your setup
Start with: Honcho only
Then, if you notice:
- “it knows me, but forgets my docs / notes / lore” → add QMD
- “I want inspectable, structured long-term knowledge” → add memory-wiki
- “I want local file-based memory consolidation I can review manually” → add OpenClaw dreaming
A very practical hybrid
For you, I think the sweet spot is:
Honcho + QMD
- Honcho = person/social memory
- QMD = local knowledge/file recall
That’s the combo I’d try before touching wiki or dreaming.
Best fully local stack
If you wanted to go mostly or fully local later:
QMD + memory-core + optional memory-wiki
- QMD for search/recall
- memory-core for local Markdown memory
- memory-wiki if you want a curated knowledge layer
That stack is more inspectable and local-first, but less naturally strong for social/user modeling than Honcho.
What I would not do
I would not start with all 4 at once.
If you run:
- Honcho
- QMD
- memory-wiki
- dreaming
…you’ll get a lot of overlap, and it becomes hard to tell which layer is actually helping versus just adding noise and maintenance.
My short ranking for your goal
For “true companion” social memory:
Honcho = huge
QMD = useful but secondary
memory-wiki = niche / later
OpenClaw dreaming = optional, probably redundant at first
If you want, I can turn this into a concrete config strategy, like:
- minimal best setup
- best hybrid setup
- maximal but sane setup
Also, if you want real-world opinions from other users, #users-helping-users is a good place to ask what stacks people actually stuck with.