feat: add cross-bank querying — multi-bank recall and reflect

[bjornslib]

Summary

• CrossBankOrchestrator for querying across multiple memory banks in parallel
• BudgetAllocator with equal/proportional/custom strategies for distributing query budget
• Reciprocal Rank Fusion (k=60) for merging and deduplicating results across banks
• Disposition reconciliation for multi-bank reflect (weighted average of bank traits)
• HTTP endpoints: POST /v1/{tenant}/cross-bank/recall and /v1/{tenant}/cross-bank/reflect
• MCP tools: cross_bank_recall, cross_bank_reflect
• Comprehensive unit tests (1400+ lines)

Motivation

Currently each bank is an isolated memory store with no way to query across banks. Many use cases (multi-agent systems, federated knowledge) need to search and reason across multiple banks simultaneously while respecting per-bank dispositions and access controls.

New files

• hindsight-api-slim/hindsight_api/engine/cross_bank.py — Orchestrator, allocator, fusion
• hindsight-api-slim/hindsight_api/api/cross_bank_models.py — Pydantic request/response models
• hindsight-api-slim/tests/test_cross_bank.py — Unit tests
• hindsight-api-slim/tests/test_mcp_tools.py — MCP tool tests

Modified files

• api/http.py — Two new cross-bank endpoints
• api/mcp.py — Register cross-bank tools
• engine/memory_engine.py — CrossBankOrchestrator property + initialization
• mcp_tools.py — Cross-bank tool implementations

Test plan

• Unit tests pass: uv run pytest tests/test_cross_bank.py tests/test_mcp_tools.py -v
• Cross-bank recall returns fused results from multiple banks
• Cross-bank reflect produces disposition-reconciled answers
• Budget allocation respects per-bank minimums
• No impact on existing single-bank operations

🤖 Generated with Claude Code

[bjornslib]

Thought that supporting multiple banks when querying was important, as some questions will stretch across banks. I hope that this works fine with the new instances concept you introduced, likely for the Hindsight Cloud version.

[nicoloboschi]

I think cross-bank operations are really off of the current model in terms of security, quality, performance and general understanding of the project.

have you tried using tags to isolate data inside a single bank?