Skip to content

Hindsight Memory Server

External Hindsight server providing semantic memory storage and reflection

Details

Hindsight is the external memory server backing the work-buddy memory subsystem (see memory). It provides storage, semantic search, retention, and LLM-powered reflection over a personal memory bank.

Server components

  • hindsight-api-slim — the HTTP server, managed via Poetry
  • hindsight-client — the Python SDK, managed via Poetry
  • External PostgreSQL with the pgvector extension (for semantic search)

Configuration

Required environment variables for the Hindsight API:

  • HINDSIGHT_API_LLM_PROVIDER — typically anthropic
  • HINDSIGHT_API_LLM_API_KEY — provider API key (Anthropic key for the default config)
  • HINDSIGHT_API_LLM_MODEL — e.g. claude-haiku-4-5-20251001

Starting the server

HINDSIGHT_API_LLM_PROVIDER=anthropic \
HINDSIGHT_API_LLM_API_KEY=$ANTHROPIC_API_KEY \
HINDSIGHT_API_LLM_MODEL=claude-haiku-4-5-20251001 \
hindsight-api

Costs

Every memory_write call costs LLM tokens (the Hindsight server runs an LLM-side embedding + summarization pass on retain). memory_read is cheap (embedding similarity only — no token cost beyond the input). memory_reflect is the most expensive — it runs LLM reasoning over the recall set.