Structured reads
Read page, row, paragraph, or line ranges after search finds the right source.
openDB
The local databasefor agent context.
openDB gives AI agents a private MCP database for files, search, and long-term memory. It keeps retrieval deterministic, local, and explainable with provenance back to the file, page, row, line range, or memory record.
Search result
Evidence from files, code, and memory.
The agent searches first, then opens only the page, row, paragraph, or line range needed to answer.
account-plan.pdfPage 6 · score 12.8
crm-export.xlsxRow 142 · ARR, owner, next step
memory:user-preferencespinned · semantic
src/parser.pylines 50-80 · grep match
1.0 Search
Make every workspace searchable before the agent opens files.
Full-text indexes over documents, code, sheets, decks, PDFs, and text let agents find the smallest useful context first. SQLite FTS5 and Postgres tsvector keep the system fast without mandatory embedding infrastructure.
Local provenance
Every result points back to a file path, source type, and retrieved slice.
CJK ready
Tokenization support keeps multilingual workspaces searchable.
Memory that can be recalled, aged, and superseded.
Semantic, episodic, and procedural memories let agents separate stable facts, dated outcomes, and reusable workflows. Pinned context surfaces immediately while older facts can decay or be replaced.
Memory docs ->prefers concise answerspinned
switched backend to SQLitesemantic
old backend preferencesuperseded
2.0 Benchmarks
Fast enough to become an agent reflex.
LongMemEval_S R@5
token savings vs command parsing
memory stress suite
embedding calls required for retrieval
Install the local context layer.
pip install opendb
opendb mcp install
opendb index /path/to/workspace