Multi-agent shared memory: best practices for two agents collaborating on the same project?

[1ain0516]

Context

I'm running two AI coding agents simultaneously:

• Agent A — general-purpose, has native Hindsight integration (hindsight_retain/recall/reflect)
• Agent B — coding-focused, no native Hindsight support, needs shell/HTTP wrapper

Both agents work on the same projects and I need them to share context: project progress, architecture decisions, debugging findings, user preferences.

What I tried

Approach 1: Three-zone bank architecture

Created three Hindsight banks on the same API instance:

• shared — user info, preferences, project config (both agents read/write)
• agent-a — Agent A engineering memory (Agent A writes, Agent B reads)
• agent-b — Agent B engineering memory (Agent B writes, Agent A reads)

Built a Python wrapper script for Agent B to call the Hindsight API via shell commands.

Problems encountered:

1. No single memory deletion — DELETE /v1/default/banks/{bank_id}/memories clears the entire bank. Can't remove individual stale/duplicate memories.
2. LLM extraction creates duplicates — DeepSeek/qwen extraction generates both Chinese and English versions of the same fact. No way to clean up without clearing+reimporting.
3. Consolidate only works within documents — POST /consolidate doesn't merge memories from different documents, so cross-document duplicates persist.
4. Wrapper script is slow — Each call requires Python process startup + HTTP + LLM inference (~0.5-1s overhead). Not great for frequent reads.
5. Small local model extracts 0 memory units — On freshly created banks, local qwen3:14b often extracts nothing. Requires expensive API models (DeepSeek) to populate.

Approach 2: CLAUDE.md + shared markdown file

Put shared context directly in a config file (auto-loaded by Agent B) and a shared markdown file in each project directory.

Problems:

• Token cost on every session start (reading the file)
• No semantic search — just linear text
• File can get stale if one agent forgets to update

Approach 3 (current): Minimal

• Agent A uses its native Hindsight memory (works great, thousands of nodes)
• Agent B reads a config file (free, auto-loaded)
• Project progress tracked in the config file
• No shared memory system between agents

Questions

1. Is there a recommended pattern for multi-agent shared memory with Hindsight? The bank isolation concept works, but the API limitations (no single delete, no cross-document consolidation) make maintenance painful.

2. Any plans for individual memory deletion? This is the biggest blocker. Without it, any LLM extraction errors or duplicates accumulate forever.

3. Any plans for cross-document consolidation? Current consolidate only merges within a single document. Multi-agent scenarios naturally create the same facts across different documents.

4. Best practice for non-native agent integration? The wrapper script works but feels hacky. Is there a lighter-weight way for agents without native Hindsight support to use it?

5. Extraction quality with small local models? qwen3:14b (local, ~14B params) often extracts 0 memory units from retains. Any tuning tips for smaller models?

Environment

• Hindsight API v0.6.0 (local, pg0 backend)
• LLM: qwen3:14b (daily, local) / DeepSeek v4 flash (consolidation, API)
• Embeddings: BAAI/bge-small-en-v1.5 (local)
• OS: Windows, consumer GPU with 8GB VRAM, 32GB RAM

[benfrank241]

@1ain0516 thanks for the detailed write-up. the multi-tool shared-memory pattern is exactly what hindsight is designed to support. The canonical version of it is one bank that every tool reads and writes through, not three banks isolated per agent. There's a recent post by a community member that walks through running that pattern across claude desktop, claude code cli, codex, openclaw, and the chatgpt apps, all pointing at one shared bank: https://hindsight.vectorize.io/blog/2026/04/07/one-memory-for-every-ai-tool.

Having multiple banks goes against the spirit of sharing information between agents and should only really be used to total isolation of tasks.