Aura Memory v1.57.0 — Crash-Safe Temporal Memory & Governed Contradictions.

Aura Memory v1.5.7 makes persistent agent memory safer, more explainable, and more resistant to stale or contradictory knowledge.

Highlights

Crash-safe temporal memory versioning

Aura now treats fact replacement as an atomic operation. The old version’s validity boundary, the new version, and their causal links are written together as one durable journal transaction.

If a process crashes during the write, Aura keeps the previous fact valid instead of leaving a broken superseded_by="pending" chain.

Legacy interrupted supersessions are automatically repaired on startup:

an existing successor is linked correctly;

if no successor was committed, the previous version is reopened;

incomplete or corrupted atomic journal tails are safely ignored.

Historical recall

Records can now carry explicit temporal validity:

valid_from

valid_until

superseded_at

The new recall_as_of() API reconstructs which version of a fact was valid at a specific business-time timestamp. Normal recall, search, and context capsules continue to return only currently valid records.

Safer contradiction resolution

Aura no longer forces every contradiction graph into two artificial sides.

A belief remains Unresolved when the conflict contains:

odd contradiction cycles such as A ↔ B ↔ C ↔ A;

multiple independent conflict components;

isolated claims that cannot be assigned to either side;

non-binary conflicts without sufficient directional evidence.

A winner is selected only when the explicit contradiction graph forms one connected bipartite component.

Governed memory promotion

Memory level controls durability—not truth.

Promotion into Domain or Identity is now blocked for records that are contradictory, unstable, temporally inactive, or lack sufficient supporting evidence. Retrieving an old rule no longer makes its evidence appear recent.

This prevents stale policies from becoming permanent simply because they were recalled frequently or had already survived for a long time.

Inspectable recall decisions

explain_recall() now shows both surfaced and suppressed evidence, including:

the selected and rejected hypotheses;

the winning hypothesis and competing score;

temporal rejection reasons;

strength and top-K admission decisions;

suppressed_by_belief_resolution;

a correlation-safe trace_id.

Records from other namespaces or tenants are never exposed in the trace, even by ID.

Persistence reliability

Critical persistence failures are no longer silently discarded in temporal supersession, standalone decay, reflection, shared import, promotion, or namespace moves.

Promotion and namespace migration update live memory only after the corresponding journal write succeeds. Multi-record namespace changes are persisted as one atomic batch.

Compatibility

Existing temporal version chains remain supported. Aura automatically migrates legacy supersession boundaries and repairs incomplete legacy pending markers during startup.

Validation

622 automated tests passed

0 failures

crash and truncated-journal recovery tested

Python 3.12 Windows wheel installation tested

persistence verified across close and reopen

release metadata verified across Rust, Python, and the changelog

Installation

pip install --upgrade aura-memory

Project: https://github.com/teolex2020/aura-memoryPyPI: https://pypi.org/project/aura-memory/

Aura Memory v1.5.6

Immutable evidence lineage, deterministic context capsules, and observable recall outcomes.

Added

• Immutable evidence lineage — SHA-256 binding between a source revision, its exact byte span, and an Aura claim, with independent verification and answer-permission gates.
• Evidence-aware research ingestion — Rust and Python APIs for findings carrying document revision, source-span integrity, verification status, and citation admission.
• Context capsules — deterministic, namespace-isolated, token-bounded hot context with selection reasons, omission counts, and stable content hashes.
• Recall/search outcome telemetry — counters for total and empty formatted recall, structured recall, tier recall, and exact search operations, with Python bindings and reset support.
• Release metadata gate — CI validation that the GitHub release tag, Rust crate, Python package, runtime version, and changelog agree.

Changed

• Evidence-aware research reports are composed only from admitted findings. Free-form synthesis is omitted until synthesis can carry claim-level lineage.
• MCP stdio, MCP HTTP, and health responses now use the package __version__ instead of stale hard-coded values.
• PyPI release metadata now links to the correct aura-memory project page.
• Repository metadata and documentation now use the canonical teolex2020/aura-memory GitHub URL.

Fixed

• Prevented a valid integrity report for one source span from authorizing a claim bound to a different span.
• Prevented blocked evidence from being reintroduced through a generated research synthesis.
• Normalized blocked and superseded metadata before context-capsule filtering.
• Included the primary formatted recall() path and cache hits in empty-recall telemetry.

v1.5.5

v1.5.5 — Learned weighted-graph topology substrate

Connection strengths that recall reinforces and maintenance decays.

Added

• Topology substrate — shared, decayable weighted graph (Topology,
  Edge, NodeId, node_id_for) with serde-backed TopologyStore.

Changed

• Recall learns connections — co-surfaced records reinforce their edge.
• Maintenance ages the topology — un-reinforced edges decay (use-it-or-lose-it).
• Causal discovery reads learned weights — prefers learned topology over
  static Record.connections. Opt-in; public API unchanged.

v1.5.4 — Autonomous Cognitive Plasticity

The biggest release since the cognitive stack launched.

AuraSDK now ships a complete autonomous learning loop: agents accumulate
structured experience from their own responses — locally, without fine-tuning,
without an LLM, with full operator audit trail.

What's new

Autonomous Cognitive Plasticity

• capture_experience() — observe model responses and extract structured events
• ingest_experience_batch() — queue experience for the maintenance pipeline
• PlasticityMode: Off / Observe / Limited / Full
• Anti-hallucination guards — generated claims capped, cannot overwrite recorded facts
• Plasticity risk scoring with auto-throttling
• purge_inference_records() / freeze_namespace_plasticity() — full operator control

Salience & Reflection

• mark_record_salience() — what matters persists longer, decays slower
• Maintenance-time reflection synthesis — bounded summaries of recurring patterns
• get_reflection_summaries() / get_latest_reflection_digest()

Contradiction Governance

• get_contradiction_clusters() — explicit grouping of conflicting evidence
• get_contradiction_review_queue() — prioritized operator review surface
• Unresolved-evidence markers in recall explanations

Production Integrity

• Concept layer now persists across restarts
• Belief reranking active by default (was Off)
• Concept partition cap — bounded maintenance cost on large corpora
• Startup validation and persistence manifest

Honest Answer Support

• Non-anthropomorphic phrasing hints for significance, uncertainty, contradiction
• Ready for agent / UI consumption via answer_support in recall explanations

Upgrade

pip install --upgrade aura-memory

v1.5.3 — Fix: ExplicitTrusted pipeline (policy hints now work)

Bug Fix: explicit causal links → policy hints

When users called link_records(cause_id, effect_id, "causal"), the SDK
accepted the links but never produced policy hints. Fixed.

What was broken

CausalEvidenceMode::ExplicitTrusted was added in 1.5.2 but the causal
and policy pipeline still applied strict repeated-window criteria,
silently rejecting all user-declared causal links.

What was fixed

• Causal evidence gates now respect explicit trust signals
• Policy seed selection accepts explicit-backed patterns
• Confidence scoring falls back to record level when beliefs are not yet formed
• Causal and policy engines now persist correctly between sessions

Result

After 2+ explicit causal links, the cognitive stack surfaces actionable
policy hints — avoid, warn, verify_first — that strengthen over time
as more evidence accumulates.

Upgrade

pip install aura-memory==1.5.3

v1.5.2 update MCP Claude

fix: MCP write bare JSON lines — remove Content-Length from responses

Claude Desktop expects bare JSON lines (not Content-Length framing).
1.5.1 was released before this fix landed — 1.5.2 corrects it.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

v1.5.1 — MCP transport fix

• Fix: MCP stdio transport rewritten — eliminated per-byte read latency
• Fix: search tool Level serialization error
• Add: smithery.yaml + mcp.json for MCP registries
• Add: Cursor/VS Code install config in README

v1.5.0 — Full Cognitive Pipeline

v1.5.0 — Full Cognitive Pipeline

What's new

• Full 5-layer cognitive recall pipeline: Belief → Concept → Causal → Policy

  • Each layer applies bounded reranking: ±5% / ±4% / ±3% / ±2%
  • Scope guards on all phases: min 4 results, zero result removal, top_k ≤ 20

• enable_full_cognitive_stack() — one call activates everything

  brain.enable_full_cognitive_stack()

v1.4.1

AuraSDK v1.4.1 is a cleanup release focused on product truth.

• deterministic local memory engine for AI agents
• production path: Record -> Belief -> bounded recall rerank
• concept / causal / policy remain advisory
• public repo cleaned from internal research artifacts
• docs and package metadata aligned with the real shipped product

Repository:
https://github.com/teolex2020/AuraSDK

v1.4.0

What's New

• Functional semantic memory types — semantic_type in store/search/MCP surfaces
• Semantic-aware recall — formatted output respects memory semantics
• Semantic-aware search — filtering and ranking by semantic type
• Phase-based insights — new semantic detectors in maintenance
• Semantic-aware maintenance — decay/consolidation adapts to semantic types
• Public docs & examples hardened — IP protection, cleaner examples

Stats

• 348 tests passing
• Backward compatible with 1.3.x integrations
• Pure Rust, zero LLM calls, sub-millisecond recall