Real-world usage report: empty representations, duplicate conclusions, no filtering — feedback from credit user

[gueysito]

Context

I've been using Honcho (SDK v2.0.1, managed service) for about 4 weeks with a single AI agent (workspace biggie, 2 peers). The agent runs daily conversations via Telegram. Dialectic budget was recently set to 1800 chars / medium reasoning.

Sharing this as constructive feedback — I'm on credits, not a paying customer yet, and I want to help improve the product. Honcho has been genuinely useful for my setup, and I'd like to see these gaps addressed so I can expand to a second agent.

What's working well

• Peer cards are excellent — rich, accurate, well-structured biographical/preference data for both user and AI peer
• Conclusion generation is capturing real preferences, rules, project context, and behavioral patterns
• Session management works cleanly (53 sessions, message storage reliable)
• The overall concept of persistent cross-session memory is exactly what I need

What's not working

1. Empty representations (relates to #524, #494)

Despite having 762 conclusions stored, calling peer.representation("other_peer") returns an empty string for both peers. This is the primary feature I'm paying dialectic budget for — the synthesis layer that should turn raw conclusions into a coherent understanding. It's been empty since activation (~4 weeks).

yuki = client.peer("yuki")
rep = yuki.representation("biggie")
print(repr(rep))  # ''

2. Duplicate conclusions (~14% of total)

54 unique strings appear multiple times across the 762 conclusions, totaling 110 duplicate entries. Examples:

• "biggie's home directory is /Users/Hermes" — 11x
• "biggie dislikes overengineering and token waste" — 11x
• "biggie wants Yuki to be a high-trust creative/technical partner" — 7x

I understand the maximalist storage philosophy (per maintainer comments on #444), but 11 copies of the same fact seems like the dreaming/consolidation system isn't keeping up.

3. No way to filter low-value conclusions

A significant portion of conclusions are infrastructure facts (file paths, B2 endpoints, log locations) that are low-value for cross-session recall. I'd love a way to guide the deriver — something like PR #430's custom instructions approach would be ideal.

What I'd find helpful

• Guidance on whether empty representations are expected with my config, or if there's something I should change
• Any ETA on deriver custom instructions (fix: implement deriver custom instructions #430) — that would solve the filtering problem
• Whether schedule_dream() would help kick-start representation generation

Environment

• SDK: honcho v2.0.1 (Python)
• Service: managed (api.honcho.dev)
• Config: 1 workspace, 2 active peers, hybrid memory mode, 1200 chars / medium reasoning
• Usage: ~4 weeks, 53 sessions, 762 conclusions