SEAM

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

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

Linux / WSL2:

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

Public release installer shape, for later release packaging:

irm https://example.com/seam/install.ps1 | iex

curl -fsSL https://example.com/seam/install.sh | sh

Those public URLs are placeholders until SEAM has a published installer host. Use the private gh repo clone commands above for this repo today.

60-Second Demo

After install, open a new terminal:

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.

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 serve exposes JSON-lines tool calls for MCP-style 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

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 index
seam reindex
seam dashboard
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
• Installer reference: installers/README.md
• Troubleshooting: docs/errors.md
• Task runbooks: docs/howto/README.md
• Active/inactive code layout: docs/CODE_LAYOUT.md

Default persistent database paths:

• Windows: %LOCALAPPDATA%\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

Benchmark evidence is the proof layer for SEAM's commercial value; it does not grant commercial, hosted, SaaS, API, managed-service, embedded, redistribution, or closed-source rights.

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

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 is proprietary source-available software under the SEAM Source-Available License. The repository is public for review, evaluation, and contribution; no open-source license is granted.

You may review, evaluate, fork for contribution, and use SEAM locally for personal, educational, research, or non-commercial purposes. You may not host, SaaS, sell, embed, redistribute, deploy for customers, commercialize, or use SEAM in a closed-source or commercial project without a separate written commercial license from the copyright holder.

Commercial use is available with permission: get a separate written commercial license from the copyright holder before using SEAM in a paid product, hosted service, customer deployment, closed-source product, or business offering.

Contributions grant the project owner rights to keep developing SEAM and to commercially license SEAM without needing later contributor permission.

See LICENSE and NOTICE.