kedger 0.2.0

Kedger — casual local-first memory for coding agents.
Git remembers what changed. Agent chats remember why… until they don’t.
Kedger keeps the why — decisions, rejects, constraints — and lets you hand it to the next agent.

Why · Team · Memory layers · Research · Privacy · Install

The problem we’re actually solving

Engineering teams now do real work inside Cursor / Claude sessions.

Then the chat ends. Context compacts. A teammate opens the same repo.

Their agent starts cold.

You already lived through:

• “Wait, why not Redis?”
• “Didn’t we reject cookie sessions?”
• “Which files did the last agent actually touch?”
• Slack paste that loses half the judgment

Code is versioned. Judgment isn’t.

That’s the gap Kedger exists for — not another living wiki of the whole codebase, and not a cloud “team brain” that syncs everything by default.

Why use this

Use Kedger if you want the next agent (you tomorrow, or a teammate today) to boot with:

1. What we decided
2. What we rejected (negative knowledge is gold)
3. What constraints still bind
4. What files / ops were in flight

…without dumping the whole transcript into git, and without putting session memory on someone else’s server.

Skip Kedger if you mainly want a repo documentation wiki that drifts with the code graph — different product, different job.

How it helps a team

Team help here is handoff, not ambient contact sharing.

Alice’s agent learns the hard parts
        │
        ▼
   Anchors + ops land in her local store
        │
        ▼
   she seals a .kxp for Bob (on purpose)
        │
        ▼
   Bob opens it → hydrate → new chat continues

So the team benefit is: judgment can move with the work, the same way a branch moves — when someone chooses to pass it.

Same person / new laptop works the same way (pack-export → hydrate --pack). No peer card required.

Why this shape (and not ambient sharing)

We deliberately chose share_mode=explicit_only.

Kedger’s bet: make capture + Anchors + sealed handoff excellent first.
Share is a sealed file (.kxp) you send like a private USB stick — Slack, Drive, USB — encrypted to the recipient’s keys.

No ambient contact graph. No “everyone in the org can read your chat memory.”
Easy continuity for people on the task. Hard/no continuity for everyone else.

Memory architecture (L0→L4)

Kedger does not treat the raw chat log as memory.
Lock: durable memory must be compact-native, not compact-rescued — anything that must survive has to live outside the transcript before the model window is destroyed.

hooks → redact → L0
              ↓ patch L1 on state deltas
              ↓ cognify / promote → L2 + L3
              ↓ seal → L4 (.kxp) → hydrate

Anchors — the atom that must survive

Under pressure we keep: constraints → rejections → decisions → goal → next_step → … → raw L0 dies first.

Ops + graph (shipped alongside Anchors)

• Ops layer: files touched, +/-, tool fails — so hydrate isn’t only philosophy.
• Graph: Anchors ↔ entities ↔ episodes (ABOUT / SUPERSEDES / SUPPORTS / …). Hydrate walks under a budget — no whole-store dump into the prompt.

Constitution: docs/OPEN_SOURCE_MEMORY_ARCHITECTURE.md.

Why this architecture (research)

We didn’t invent layers for branding. The stack is a research → measure → refine loop over agent-memory, compaction, graph, privacy, and seal literature.

Corpus (honest numbers): Track 0 keeps a 500-slot FULL deep-read ledger (~484 distinct arXiv primaries + eng/crypto FULL through Batch25), plus survey-indexed maps and eval harnesses — not keyword theater. Inventory: docs/research/CORPUS_INVENTORY.md · matrix: docs/research/KEDGER_STAGE_RESEARCH_MATRIX.md · queue: docs/research/queue/FULL_QUEUE_500.md.

Influences called out in the constitution include MemGPT, Mem0, Graphiti/Zep, HippoRAG, GraphRAG/LightRAG, Nemori, Cognee, Generative Agents, and the privacy/seal cluster — then measured in tests/eval/ (strict handoff, spectrum insight, share probes).

Privacy tradeoff

Defaults that are true today

• Store lives under ~/.kedger/ (not the source of truth in git)
• Redact-before-persist on ingest
• Share only when you seal + send a .kxp
• Peer card = public keys only (safe to Slack)
• Unauthorized open fails closed (not found) — no “yes that pack exists, here’s a peek”

Tradeoff we accept

• You must actively send a pack (friction by design)
• Revoke + reseal stops new packs; old offline .kxp files stay readable to old keys (crypto reality — we say it out loud)
• Full DB at-rest encryption is later (docs/PHASE_F_DEFERRED.md)

If you didn’t send a pack, they don’t get your session. That’s the product.

What it looks like

kedger remember reject "no Redis — use Postgres"
kedger cognify --force --promote
kedger hydrate --live
# → what the next agent will see

Install (60 seconds)

pip install "kedger>=0.2.0"
cd /path/to/your-app
kedger init --name alice

init writes keys, repo policy, and Cursor / Claude hook packs. Trust the workspace once, start a new chat, keep working. Kedger captures in the background.

After upgrade to 0.2.0: re-run kedger hooks install in each app repo, then verify inject with docs/PROMPT_INJECT_VERIFY.md. For peer handoff dogfood, see docs/PEER_TRIALS.md or bash scripts/peer_trial.sh.

Two people, two agents

Alice                                         Bob
─────                                         ───
kedger init --name alice                      kedger init --name bob
…agent works…                                 kedger peer card --out bob.kedger.json
                                         ◄──── card (public keys only)
kedger peer send --to bob.kedger.json --out-dir ./xfer
────── send ./xfer/*.kxp ─────────────────►
                                              kedger peer open hf_….kxp
                                              kedger hydrate --live
                                              → new IDE chat with Alice’s Anchors

CLI

kedger doctor                 # health + locks (share_mode=explicit_only)
kedger remember reject "…"    # durable Anchor
kedger cognify --force --promote
kedger hydrate --live         # next-agent preview
kedger peer card|send|open    # person-to-person sealed handoff

Product locks

Shipped: hooks, claim extract, Anchors + ops, zlib transcript, sealed packs, peer card/send/open, prompt-time inject, MCP pull, consolidate, seed-IDF retrieve.
Not yet: LLM-every-turn distill, sync service, at-rest DB encryption — docs/PHASE_F_DEFERRED.md.
Proof: Beta — mechanically tested (CI + strict handoff evals). Human peer trials still pending — docs/PEER_TRIALS.md. Not a published user study.
Verify inject: docs/PROMPT_INJECT_VERIFY.md (SessionStart may drop; per-prompt inject is the reliable path).

Contributing

• Peer break? → issue template
• CONTRIBUTING.md · SECURITY.md · docs/MARKETING.md
• Architecture constitution: docs/OPEN_SOURCE_MEMORY_ARCHITECTURE.md

pip install -e ".[dev]"
pytest -q
./scripts/smoke_transfer.sh && ./scripts/smoke_wheel_install.sh && ./scripts/smoke_peer_handoff.sh

License

Apache-2.0 — LICENSE.