Seam Runtime

Give your local AI agent persistent memory in one command.

SEAM is a local memory runtime for agents. It stores durable MIRL records in SQLite, retrieves compact context with lexical, graph, temporal, and vector signals, tracks provenance, exposes a dashboard/API, and gates benchmark claims before they are treated as real progress.

Install

GitHub package install for the SEAM runtime:

python -m pip install "seam-runtime @ git+https://github.com/BlackhatShiftey/Seam_Runtime.git@main"

Install with REST API and dashboard extras:

python -m pip install "seam-runtime[server,dash] @ git+https://github.com/BlackhatShiftey/Seam_Runtime.git@main"

Once release tags exist, replace @main with a pinned tag such as @v0.1.0. The clone-and-installer flows below remain the full operator setup path for repo-local development, persistent state setup, and platform shims.

Private repo install requires an authenticated GitHub CLI session.

Windows PowerShell:

gh repo clone BlackhatShiftey/Seam Seam; cd Seam; powershell -ExecutionPolicy Bypass -File .\installers\install_seam_windows.ps1

macOS:

gh repo clone BlackhatShiftey/Seam Seam && cd Seam && sh ./installers/install_seam_macos.sh

Linux / WSL2:

gh repo clone BlackhatShiftey/Seam Seam && cd Seam && sh ./installers/install_seam_linux.sh

Repo-local Linux development bootstrap:

gh repo clone BlackhatShiftey/Seam Seam && cd Seam && sh ./installers/install_seam_linux.sh --dev

Agent Setup Prompt

After cloning SEAM, paste this prompt into your coding agent to have it install SEAM, verify the local setup, and configure SEAM as persistent memory for the workspace.

You are setting up SEAM from this repository.

Goal:
Install SEAM completely for local development and operator use, then configure
it as persistent memory for this agent/workspace.

Rules:
- Read `AGENTS.md` first and follow repo-local instructions.
- Do not expose, print, copy, delete, or summarize secrets.
- Do not ingest secrets, `.env` files, credential files, private keys, provider
  session links, ignored local artifacts, or private chat/share links.
- API keys, local `.env` files, and local `.conf` files are operator-owned.
  The operator can set them in the SEAM Web UI Settings panel or maintain them
  manually in ignored local config files.
- Prefer project installers and documented commands over ad hoc setup.
- Do not install `bench-judge`, `bench-mem0`, or `bench-zep` unless the operator
  explicitly approves provider/API-key benchmark dependencies.
- If a command fails, stop and report the exact command and error. Check
  `docs/errors.md` for the symptom before giving up, and report whether that
  reference resolved it.

Steps:
1. Confirm the current directory is the SEAM repo.
2. Run the platform installer:
   - Linux/WSL2: `sh ./installers/install_seam_linux.sh --dev`
   - macOS: `sh ./installers/install_seam_macos.sh --dev`
   - Windows PowerShell: `powershell -ExecutionPolicy Bypass -File .\installers\install_seam_windows.ps1 -Dev`
3. Install useful local extras for normal operator work:
   `python -m pip install -e ".[server,dash,pgvector,sbert,rerank]"`
4. Verify the install:
   `seam doctor`
5. Ask the operator to set any needed provider keys and local config before
   enabling paid/provider-backed features:
   - Web UI path: run `seam webui`, open Settings, enter provider keys,
     `SEAM_CHAT_API_KEY`, `SEAM_CHAT_BASE_URL`, `SEAM_CHAT_MODEL`,
     `SEAM_PGVECTOR_DSN`, `SEAM_API_TOKEN`, or `SEAM_LOCAL_ENV` as needed, then
     save the local env from the Settings panel.
   - Manual path: create or edit ignored local `.env` or `.conf` files, export
     the needed variables in the shell, and never commit or ingest those files.
6. Re-run:
   `seam doctor`
7. Ingest safe repo context as persistent memory:
   `seam ingest README.md --persist`
   `seam ingest AGENTS.md --persist`
   `seam ingest PROJECT_STATUS.md --persist`
   `seam ingest REPO_LEDGER.md --persist`
8. Test memory retrieval against the ingested docs:
   `seam memory search "current SEAM repo status"`
   `seam context "current SEAM repo status" --retrieval-mode mix --view prompt`
9. Prove write-then-read persistence with a fact this session creates, not a
   pre-existing doc:
   `seam remember "SEAM setup smoke test <unique token>"`
   `seam memory search "setup smoke test"`
   Confirm the exact fact just written comes back before treating memory as
   working.
10. If this agent supports MCP, configure it to launch:
   `seam-mcp`
   Or, when pgvector is needed and Docker is available:
   `seam-mcp --ensure-pgvector`
   Verify the server actually responds: issue an MCP tool-list/discovery call
   against the running process and confirm SEAM's tools (memory search/get,
   context, ingest) appear before reporting MCP as configured.
11. Report back with:
   - install path used
   - optional extras installed
   - whether `seam doctor` passed
   - whether API keys/local config were set in Web UI Settings or manually
   - whether memory search/context returned useful repo context
   - whether the write-then-read smoke test round-tripped correctly
   - whether MCP was configured and its tool list verified, or only CLI
     memory is available

60-Second Demo

After install, open a new terminal. The same commands work on Windows PowerShell and Linux / WSL2 because seam is a platform-agnostic shim:

seam doctor
seam ingest README.md --persist
seam memory search "persistent agent memory"
seam retrieve "persistent agent memory" --mode mix --budget 5
seam context "persistent agent memory" --retrieval-mode mix --view prompt
seam dashboard --snapshot --no-clear

Inside the dashboard, use reload or /reload to refresh the visible runtime state, metrics, panels, and chart surfaces without restarting.

To configure API keys and local runtime settings without editing files by hand, run the browser Web UI and open Settings:

seam webui --host 127.0.0.1 --port 8765

Settings covers provider keys, chat/API settings, embedding settings, database paths, pgvector DSNs, SEAM_LOCAL_ENV, REST API tokens, and save/reload local env controls. Operators can also maintain ignored local .env or .conf files manually; those files must not be committed or ingested as memory.

Why SEAM

• Persistent local memory: SQLite is the canonical source of truth.
• Efficient RAG: vector, graph, hybrid, and mix retrieval modes.
• Progressive disclosure: seam memory search gives compact IDs first; seam memory get <ids> fetches full records only when needed.
• Agent bridge: seam mcp stdio / seam-mcp exposes a standard MCP server for Gemini, Claude, Cursor, and other agents. Gemini's project config starts it with --ensure-pgvector so Docker Compose pgvector is ready before MCP tool discovery. seam mcp serve remains available for legacy JSON-lines wrappers.
• Provenance: records keep refs, evidence, trace edges, and source document status.
• Benchmark discipline: benchmark bundles are hash-verified, diffed, gated, and separated from holdout publication runs.
• Operator surface: CLI, Textual dashboard, REST API, and installer shims all use the same runtime.

Core Commands

Cross-platform (Windows PowerShell and Linux / WSL2 share the seam shim):

seam ingest path/to/file.txt --persist
seam remember "SEAM stores durable memory for agents."
seam memory search "durable memory"
seam memory get clm:1,sta:ent:project:seam --timeline
seam retrieve "durable memory" --mode mix --trace
seam context "durable memory" --retrieval-mode mix --view evidence
seam surface compile path/to/file.txt --output file.seam.png --mode rgb24
seam surface query file.seam.png "durable memory"
seam shell
seam index
seam reindex
seam dashboard
seam mcp stdio
seam-mcp --ensure-pgvector
seam mcp serve
seam serve --host 127.0.0.1 --port 8765
seam benchmark run all --persist
seam benchmark gate seam-benchmark-report.json

RAG Architecture

SEAM takes the useful parts of graph RAG systems while keeping canonical memory inside MIRL:

1. ingest text or files
2. compile semantic records into MIRL
3. persist canonical records and document status in SQLite
4. derive vector indexes and graph edges from record IDs
5. retrieve with vector, graph, hybrid, or mix
6. build a token-bounded PACK for the agent

Vector stores are acceleration layers, not source of truth. PgVector and the SQLite vector table can be rebuilt from MIRL records.

See docs/RAG_ARCHITECTURE.md for the LightRAG and claude-mem comparison, retrieval mode details, and agent bridge notes.

Setup Details

• Command cookbook: docs/setup.md
• Fresh Linux resume checks: docs/setup.md#resume-current-repo-state-on-fresh-linux
• Installer reference: installers/README.md
• Troubleshooting: docs/errors.md
• Task runbooks: docs/howto/README.md
• Active/inactive code layout: docs/CODE_LAYOUT.md

Operator Manual

For help beyond the quickstart, use these docs as the operator manual:

• Operator guide - day-to-day commands, doctor checks, benchmark posture, and failure triage (Windows, macOS, and Linux).
• macOS guide - install paths, Application Support layout, Docker/pgvector, MCP, and macOS troubleshooting.
• Setup guide - platform setup, installer flows, dashboard chat model configuration, and supported command shapes.
• Task runbooks - short workflows for common operator tasks.
• Engineering manual - architecture, security, change/test/incident SOPs, and verification discipline.
• Troubleshooting and error index - look up failures by symptom or error type before changing code.

Error Index

Start with docs/errors.md. Current indexed failure types include:

• ModuleNotFoundError: No module named 'textual'
• SEAM doctor: FAIL
• PgVector: configured but unreachable
• Chroma path/index sync failure
• Benchmark bundle verification failure
• HTTP 429 provider quota or rate-limit symptoms

Default persistent database paths:

• Windows: %LOCALAPPDATA%\SEAM\state\seam.db
• macOS: ~/Library/Application Support/SEAM/state/seam.db
• Linux / WSL2: ~/.local/share/seam/state/seam.db

Optional Extras

python -m pip install -e ".[dash]"
python -m pip install -e ".[server]"
python -m pip install -e ".[pgvector]"
python -m pip install -e ".[sbert]"
python -m pip install -e ".[agent]"
python -m pip install -e ".[rerank]"
python -m pip install -e ".[all-extras]"

Extras keep the base install focused:

• dash: Textual dashboard
• server: FastAPI/Uvicorn REST API
• pgvector: PostgreSQL PgVector adapter
• sbert: local sentence-transformer embeddings
• agent: reserved agent bridge extra; current stdio bridge has no extra dependency
• rerank: reranker model dependencies

REST API

Install the server extra:

python -m pip install -e ".[server]"

Run locally:

seam serve --host 127.0.0.1 --port 8765

Useful endpoints:

• GET /health
• GET /stats
• POST /compile
• POST /compile-dsl
• GET /search?query=durable+memory&budget=5
• POST /context
• POST /lossless-compress
• POST /persist

Set SEAM_API_TOKEN to require Authorization: Bearer <local-token> for protected endpoints.

Benchmark Glassbox

seam benchmark run all --persist --output seam-benchmark-report.json
seam benchmark show latest
seam benchmark verify seam-benchmark-report.json
seam benchmark gate seam-benchmark-report.json
seam benchmark diff <baseline-report.json> seam-benchmark-report.json

Measure Progress (Or Regression)

The visual-memory loop is a measurable iteration engine. To prove a change improves SEAM rather than regressing it:

# 1. capture baseline
seam benchmark run all --persist --output baseline.json

# 2. make the change

# 3. capture after-state and compare
seam benchmark run all --persist --output after.json
seam benchmark diff baseline.json after.json
seam benchmark gate after.json

benchmark diff shows per-case green/red/gray deltas and added/removed cases. benchmark gate enforces the release-blocking minimums across all eight families. To extend coverage of structured document features, add a fixture case to benchmarks/fixtures/surface_cases.json; if the underlying extractor does not exist yet in seam_runtime/lossless.py:_structural_quote_spans, the gate fails and the fix is local. See docs/howto/README.md section 4 for the failing-case-driven extension runbook.

Publication Discipline

Benchmark evidence is the proof layer for SEAM's commercial value. It does not grant trademark rights, imply endorsement, or provide access to private hosted, enterprise, customer-specific, or unreleased SEAM offerings.

Publication claims must include bundle hash, case hashes, fixture hashes, git SHA, diff output, gate output, and holdout output when the claim is external. Do not claim "best", "production proven", or "commercial-grade" unless the benchmark bundle supports that exact claim.

Machine-First Layer

The product entrypoint is simple: install SEAM, persist memory, retrieve context. Under that surface, SEAM is still machine-first:

• MIRL: canonical memory IR
• PACK: prompt-time context view
• SEAM-LX/1: exact machine-text envelope for lossless workflows
• SEAM-RC/1: directly readable compressed machine language
• SEAM-HS/1: lossless PNG-backed surface for MIRL, RC/1, LX/1, or raw bytes

The design stance is unchanged: SQLite is canonical, derived indexes are rebuildable, lossless claims require exact reconstruction, and compressed artifacts must remain useful to an agent without hiding provenance.

License

SEAM Runtime's public core is licensed under the Apache License 2.0. The public core repository is:

https://github.com/BlackhatShiftey/Seam_Runtime

Apache-2.0 permits use, modification, redistribution, and commercial use of the public core under the license terms. It does not grant SEAM trademark or branding rights, and it does not grant access to private repositories, hosted services, enterprise modules, private connectors, private benchmark holdouts, customer-specific integrations, or unreleased methods.

Separate commercial offerings may cover hosted SEAM services, enterprise connectors, managed deployments, support, warranty, indemnity, private modules, or customer-specific implementation work.

See LICENSE, NOTICE, and COMMERCIAL_LICENSE.md.