Skip to content

Changelog

Current release: 0.11.1. Source of truth for the running version is VERSION in core/constants.py. Packed notes live in releases/manifest.json.

The file the docs and the site offer is https://indb.tech/releases/indb-latest.tar.gz. That name is this cut until the next pack. 1.0.0 has not been cut.

0.11.1 — Inhale looks sideways

Date: 2026-11-03

A new event is scored against hot events that share a token. The rest of the life is not walked. Embeddings, when present, still see the whole hot tier. Fusion no longer prints a debug line on every comparison. Replacing a fused body uses the id map.

Latest: https://indb.tech/releases/indb-latest.tar.gz
Archive: https://indb.tech/releases/indb-0.11.1.tar.gz
Self-host: https://docs.indb.tech/SELFHOST/

0.11.0 — Edges fade

Date: 2026-11-03

A graph edge that touches a weak sleeper is marked stale and kept. It no longer wakes anyone and no longer answers Echo. A sleeper that still carries mass holds its edge past the seven-day clock. Token overlap can still wake what the edge has faded.

0.10.0 — Cold stays cold

Date: 2026-11-03

A sleeper is the encrypted file, not a second copy of the life in RAM. The process keeps the id and the tokens, enough to notice a strong seed. The body is read from cold.bin when that seed wakes it, or when the id is asked for by name.

0.9.1 — Direct recall

Date: 2026-11-03

Hot memory keeps an id map. get_event, get_event_by_id, and the Raft commit wait resolve one event without walking the stream. A sleeping event is still found by id in the cold tier.

0.9.0 — Sleep & Wake

Date: 2026-11-03

  • A full hot tier exhales low-salience events into encrypted cold.bin. They are not deleted.
  • Signed events stay hot. The recent tail stays hot.
  • A seed whose token overlap or stored graph edge reaches the wake threshold pulls the sleeper back, re-indexes it, and reinforces fusion_count.
  • /api/v2/health reports hot_events, cold_events, and wake_count (wakes since process start).
  • Reference: docs/MEMORY_LIFECYCLE.md

0.8.0 — Psycho & Grain

Date: 2026-09-22
Image: indb:0.8.0
Git: e3e5882

Psycho

Pathfinding under weight ablation. Echo radiates; Psycho picks one route and prices what the observer's values did to it.

cost(a → b) = -log R(a, b) + barrier(b)

Minimising the sum of -log R maximises the product of resonances, so the cheapest path is the most plausible chain. barrier comes from the negative entries of the caller's token_weights.

The search runs twice — valuation honoured, valuation removed — and both routes are priced on the same scale.

  • GET /api/v2/psycho/profiles
  • POST /api/v2/psycho/path with from, to, profile, compare_to, token_weights
  • Headline fields: detour_cost, barrier_walked_into, identical_route
  • Profiles include psychopath (affect flat, distress ignored) and empath; the baseline is neutral
  • Reference: docs/PSYCHO.md

Grain

Identity verification in the kernel, from barcode-exact to a SynthID-style fingerprint.

Tier Analog Tolerance
exact EAN-13 / Ed25519 ~0–5%
redundant QR-style N-of-M ~30%
resonant Echo × source reputation ~30–50%
statistical n-gram fingerprint p-value
  • GET /api/v2/grain/tiers
  • POST /api/v2/grain/verify
  • Reference: docs/GRAIN.md

Auth

  • AUTH_MODE, RBAC roles, API keys
  • Anonymous writes are rejected

Product image

indb:0.8.0 ships the API, cognitive modules, WUI, and TUI. The landing site (site/) is a separate application.

0.7.0 — Origin layer

Date: 2026-01-12
Image: indb:0.7.0

  • Source entity (Sensor, AI, Human, …) and a kernel source registry
  • Reputation-weighted fusion
  • Ed25519 event signing and provenance (source_id)
  • Prism perception paradox: perception_gap, alternative_readings, observer token_weights
  • Signed memory recovery by public key_id

Earlier milestones

These shipped before the packed catalog. They are in the 0.8.0 engine. They are not later version numbers.

0.6.0 — Binary stream

  • gRPC IngestStream
  • binary_payload on events
  • JIT renderer decoupled from storage

0.5.0 — Fortress

  • Audit log
  • RBAC
  • AES-256-GCM at rest, TLS in transit
  • No external database: indb.bin and msgpack

0.4.0 — Raft

  • Leader election and log replication
  • Writes only on the leader

0.3.0 — Cognitive core

  • Event stream
  • Fusion
  • JIT interpretation
  • HTTP, WebSocket, UDP

What was not a release

An earlier draft of this file called the current release 1.0.0 and assigned AES, caching, and a 100k events/sec benchmark to that number. 1.0.0 was never cut. Cluster throughput measured for the 3-node Raft setup remains 19–28 events/sec. Optional sentence-transformers embeddings fall back to token overlap when the model is not installed.