Hindsight: AI Long-Term Memory

Hindsight is a PostgreSQL-native long-term memory service for AI agents.

Pigsty provides the app/hindsight configuration template (conf/app/hindsight.yml). By default, it uses Pigsty-managed PostgreSQL instead of Hindsight’s built-in development database.

Quick Start

curl -fsSL https://repo.pigsty.io/get | bash; cd ~/pigsty
./bootstrap
./configure -c app/hindsight
vi pigsty.yml                 # edit domain, database password, and LLM settings
./deploy.yml
./docker.yml
./app.yml

Default access URLs:

• UI: http://hs.pigsty or http://<IP>:9999
• API: http://hs-api.pigsty or http://<IP>:8888

Key Settings

conf/app/hindsight.yml overrides /opt/hindsight/.env through apps.hindsight.conf. Key parameters:

• HINDSIGHT_API_PUBLISH_PORT: public API port, default 8888
• HINDSIGHT_UI_PUBLISH_PORT: public UI port, default 9999
• HINDSIGHT_DB_HOST / HINDSIGHT_DB_PORT / HINDSIGHT_DB_NAME / HINDSIGHT_DB_USER / HINDSIGHT_DB_PASSWORD
• HINDSIGHT_API_VECTOR_EXTENSION: default pgvector
• HINDSIGHT_API_TEXT_SEARCH_EXTENSION: default native
• HINDSIGHT_API_LLM_PROVIDER: default none
• HF_HUB_OFFLINE: defaults to 1 to use bundled models; set it to 0 and provide outbound network access to download uncached models

With the default LLM provider, none, Hindsight can retain text and recall it using local embedding/reranker models. Fact extraction, reflection, and consolidation require Ollama or an OpenAI-compatible API.

The template preserves bundled and downloaded models in the hindsight-cache named volume. The database needs vector, plus pg_trgm installed in public. Check the pg_trgm location when reusing a database from another template; CREATE EXTENSION IF NOT EXISTS does not move an existing extension.

Operations

cd /opt/hindsight
make up
make log
make info
make down
make pull

References

• Hindsight project: https://github.com/vectorize-io/hindsight
• Pigsty template: https://github.com/pgsty/pigsty/blob/main/conf/app/hindsight.yml