Mneno

Mneno is a Python SDK for developers building copilots, support agents, internal AI tools, and long-running AI applications.

It is not just a memory store. Mneno is an anti-context-rot memory runtime: it keeps context useful, compact, explainable, and verifiable as an application runs over time.

Why Context Rot Matters

Long-running AI systems accumulate stale facts, duplicated notes, noisy transcripts, and vague summaries. That context eventually makes retrieval worse, not better. Mneno is being built around scoring, compaction, and explainable diffs so developers can see what memory was kept, merged, discarded, and why.

Quickstart

pip install --pre mneno

The current alpha release is 0.3.0a3.

For local development:

scripts/setup_dev.sh

Example Usage

from mneno import MemoryClient

client = MemoryClient()
client.add(
    "The user is building Mneno, a Python SDK for explainable AI memory.",
    memory_type="semantic",
    importance=0.9,
    tags=["project", "mneno"],
)

results = client.search("What is the user building?")

for result in results:
    print(result.memory.content)
    print(result.score.total)
    print(result.score.reasons)

Core Concepts

• Memory scoring: rank memories by recency, relevance, importance, and freshness signals.
• Explainable retrieval: search returns ranked MemorySearchResult objects with score components and reasons.
• Auto-compaction: preserve key claims when context gets too large or noisy.
• Explainable memory diff: show what was kept, merged, discarded, and why.
• Lifecycle and hierarchy: preserve audit history across conflicts, supersession, archival, and layer transitions.
• Sessions and timelines: organize temporal continuity and reconstruct deterministic event order.
• Local observability: inspect memory decisions with optional in-memory traces.
• Evaluation hooks: produce deterministic metrics and versioned exports for Mneno Bench.

Auto-Compaction With Explainable Diff

Mneno can preview or apply deterministic local compaction. The engine keeps preserved memory types, merges simple duplicates, discards stale low-value memories, and returns an inspectable diff. An optional LLMProvider may improve wording only after deterministic policy has selected a merge group.

from mneno import MemoryClient
from mneno.compaction import CompactionPolicy

client = MemoryClient()

client.add("User prefers Python.", memory_type="preference", importance=0.9)
client.add("User prefers Python.", memory_type="semantic", importance=0.7)
client.add("Small talk from old session.", importance=0.1)

diff = client.preview_compaction()

print(diff.summary)
print(diff.stats)

diff = client.compact(policy=CompactionPolicy())

for decision in diff.discarded:
    print(decision.reason)

Use preview_compaction() to inspect decisions without mutating storage. Use compact() to apply the diff to the local in-memory store.

Build Explainable Context

Mneno can build a deterministic context package for prompt injection. It retrieves local memories, scores them, fits them inside an approximate token budget, and explains what was included or excluded.

from mneno import MemoryClient

client = MemoryClient()
client.add("User is building Mneno.", importance=0.9)
client.add("User prefers Python 3.11.", importance=0.8)

context = client.build_context("What is the user building?", budget=50)

print(context.text)

for item in context.included:
    print(item.reason)

The current budget uses a local approximate token estimator based on whitespace splitting. No tokenizer, LLM, embedding, or external service is required.

Policy-Driven Context Building

Use presets for common cost and recall profiles:

context = client.build_context("What matters now?", preset="cheap")
context = client.build_context("What matters now?", preset="high_recall")
context = client.build_context("What is the current agent state?", preset="agent_state")

Use a custom policy when you need explicit control:

from mneno.context import ContextPolicy

custom = ContextPolicy(
    max_tokens=800,
    reserve_tokens=100,
    min_score=0.25,
    strategy="importance",
)

context = client.build_context("What matters now?", policy=custom)

Available presets:

• cheap: smallest useful context, optimized for cost.
• balanced: default quality/cost tradeoff.
• high_recall: includes more context to avoid missing useful memories.
• agent_state: prioritizes operational state, goals, constraints, and preferences.

Persistent Storage

Mneno defaults to in-memory storage for tests, demos, and short-lived processes. For local persistence, use JSON file or SQLite storage.

from mneno import MemoryClient
from mneno.storage import JSONFileStorage, SQLiteStorage

client = MemoryClient(storage=JSONFileStorage("data/memories.json"))
client.add("User prefers Python.", importance=0.9)

client = MemoryClient(storage=SQLiteStorage("data/mneno.db"))
client.add("User is building Mneno.", importance=0.95)

Use:

• InMemoryStorage for tests and demos.
• JSONFileStorage for simple local development and human-readable files.
• SQLiteStorage for durable local apps and larger memory sets.

External database adapters will come later as optional integrations.

Import, Export, Backup, And Restore

Mneno exports memories with a versioned JSON format. Backups are normal export files, and import validation prevents silent corruption from unknown formats or unsupported versions.

from mneno import MemoryClient

client = MemoryClient()
client.add("User prefers Python.", importance=0.9)

payload = client.export_json()

client.export_json("exports/memories.json")

backup_path = client.backup()

client.restore("backups/mneno-backup-20260101-120000.json")

client.import_json("exports/memories.json", mode="skip_existing")

Import modes:

• append: add imported memories; conflicting IDs are copied with new IDs.
• replace: clear current storage first, then import.
• skip_existing: keep existing memories with matching IDs.
• overwrite: replace existing memories with imported memories of the same ID.

Cloud sync is future work. The current import/export tools are local and deterministic.

Conflict Detection And Memory Resolution

Mneno detects simple contradictions, duplicates, and superseding facts with deterministic local heuristics. It never deletes memories automatically. Resolution updates lifecycle metadata, records audit events, and keeps conflict reports explainable.

from mneno import MemoryClient

client = MemoryClient()

client.add("User prefers Python 3.10.", memory_type="preference")

result = client.add_with_report(
    "User now prefers Python 3.11.",
    memory_type="preference",
)

for report in result.conflict_reports:
    print(report.conflict_type, report.reason)

active = client.search("What Python version does the user prefer?")

Memory statuses are active, superseded, archived, and conflicted. Superseded and archived memories are excluded from default retrieval and context building; pass include_inactive=True to inspect them. Audit history preserves why a memory changed status, which memory caused the change, and the evidence behind the report.

Hierarchical Memory Organization

Mneno organizes memories into cognitive layers so short-lived notes, working state, durable knowledge, and operational instructions can age and retrieve differently.

Layers:

• short_term: transient session memories with low durability.
• working: current task state, immediate goals, and active constraints.
• episodic: session or project history and temporal events.
• semantic: durable factual knowledge and stable user/project information.
• operational: instructions, constraints, system state, and active workflows.
• archived: inactive historical memories, preserved but excluded by default.

from mneno import MemoryClient

client = MemoryClient()

client.add(
    "Current task: build benchmark UI.",
    memory_type="operational",
)

client.evaluate_hierarchy()

operational = client.list_by_layer("operational")

Retrieval is lifecycle-aware: operational, working, and semantic memories receive priority; archived and superseded memories are excluded unless you pass include_archived=True or include_inactive=True. Frequently useful memories can be promoted, stale temporary memories can be archived, and every transition is recorded in audit history. Mneno archives instead of silently deleting.

Sessions And Timelines

Sessions group memories into temporal units so long-running applications can reconstruct what happened in one session or across multiple sessions. A memory belongs to one primary session for now; richer graph-style linking is future work.

from mneno import MemoryClient

client = MemoryClient()

session = client.create_session(
    title="Mneno benchmark platform work",
)

client.add(
    "Implemented Step 14 timeline support.",
    session_id=session.id,
)

timeline = client.build_timeline(
    session_ids=[session.id],
)

for event in timeline.events:
    print(event.content)

Session-aware retrieval gives the active session a bounded deterministic boost and gives related sessions a smaller continuity boost. Query relevance remains able to outrank irrelevant current-session content. Lightweight lexical matching normalizes case, punctuation, stopwords, and simple plural variants while rewarding short phrase matches without counting repeated keywords. Use get_session_context() to build context for one session, find_related_sessions() to recover continuity across related sessions, and build_timeline() to reconstruct deterministic event order from timestamps and per-session sequence indexes. Sessions, timeline metadata, and memory session links are preserved by JSON storage, SQLite storage, import/export, backup, and restore.

Observability And Tracing

Mneno includes optional local tracing for debugging retrieval, context building, compaction, hierarchy transitions, conflicts, extraction, sessions, and timelines. Tracing is disabled by default, uses only in-memory storage, and does not send events to external services.

from mneno import MemoryClient

client = MemoryClient(trace_enabled=True)

client.add("User prefers Python.")
client.search("Python")

trace = client.get_trace(client.last_trace_id)

for event in trace.events:
    print(event.event_type, event.message)

Use client.list_traces() and client.clear_traces() to inspect or reset local traces. TraceInspector can summarize traces, filter events by memory, session, or event type, explain a memory decision, and export a stable JSON-compatible dictionary. Trace data stays local and structured secret fields are redacted before recording.

Retrieval and context traces include per-candidate score components, final rank, inclusion or exclusion decisions, session and continuity boosts, and stable source identity fields such as source_id, dataset_id, locomo_id, and original_id when those values are present in memory metadata. Benchmark adapters can also use dataset_memory_id.

Evaluation And Benchmark Integration

Mneno exposes local evaluation hooks for future Mneno Bench adapters. The SDK does not include benchmark datasets, online telemetry, or external evaluation dependencies; it provides stable report models, deterministic structural metrics, trace export, and benchmark payload helpers.

from mneno import MemoryClient

client = MemoryClient(trace_enabled=True)

client.add("The user is building Mneno.", importance=0.9)

report = client.evaluate_context(
    query="What is the user building?",
)

print(report.metrics)

Evaluation helpers cover retrieval precision/recall/MRR, token efficiency, context utilization, compaction reduction, latency, scanned/selected memory counts, and explainability coverage. evaluate_search(), evaluate_context(), and evaluate_compaction() return typed, serializable operation results with to_dict() and to_json() helpers.

EvaluationReport groups metrics and trace references for one benchmark run. export_benchmark_result() produces the stable Mneno Bench payload:

{
  "format": "mneno.benchmark.result",
  "version": 1,
  "benchmark": "synthetic",
  "created_at": "2026-06-07T12:00:00Z",
  "metrics": [],
  "traces": [],
  "metadata": {}
}

Trace exports use a separate versioned envelope:

{
  "format": "mneno.trace",
  "version": 1,
  "trace": {}
}

Mneno Bench can implement BenchmarkAdapter, run SDK evaluation wrappers, collect their metrics and trace IDs, and persist these envelopes for comparison across datasets and runtime versions. External LOCOMO, LongMemEval, and BEAM adapters remain outside the core SDK.

Alpha releases are published to PyPI from version tags through GitHub Actions and PyPI Trusted Publishing. Install the latest pre-release with:

pip install --pre mneno

Provider Architecture

Mneno is provider-agnostic. The core runtime remains deterministic and local by default, while Phase 2 introduces stable plugin contracts for future integrations.

Current provider interfaces:

• EmbeddingProvider
• LLMProvider
• RerankerProvider
• ProviderRegistry

No real OpenAI, Anthropic, Cohere, HuggingFace, vector database, graph database, or network provider is included in core. Future integrations should be optional adapters that implement these contracts.

from mneno.providers import ProviderRegistry
from mneno.providers.embedding import DummyEmbeddingProvider

registry = ProviderRegistry()
registry.register_embedding("dummy", DummyEmbeddingProvider())

provider = registry.get_embedding("dummy")
vectors = provider.embed(["User prefers Python."])

The dummy providers are deterministic and local-only. They exist for tests, examples, and adapter development.

Semantic Retrieval With Providers

Mneno can optionally use an EmbeddingProvider to add semantic relevance to search and context building. The default behavior remains deterministic keyword/scoring retrieval when no provider is configured.

No real external providers are included in core yet. Future integrations can implement the EmbeddingProvider protocol without changing Mneno's local runtime.

from mneno import MemoryClient
from mneno.providers.embedding import DummyEmbeddingProvider

client = MemoryClient(
    embedding_provider=DummyEmbeddingProvider(),
)

client.add("User is building Mneno, an SDK for explainable AI memory.", importance=0.9)
client.add("User likes Italian food.", importance=0.5)

results = client.search(
    "AI memory SDK project",
    use_semantic=True,
)

for result in results:
    print(result.memory.content)
    print(result.score.total)
    print(result.score.reasons)

Use use_semantic=False to force deterministic fallback search even when an embedding provider is configured. Mneno computes embeddings at search time for now; embedding caches, vector indexes, and vector database adapters are future work.

Reranker-Powered Retrieval

Mneno can optionally apply a second-stage reranker after candidate generation. Candidate generation still uses the deterministic scorer and, when configured, optional semantic relevance. The reranker only reorders candidates and adds explainable reranking metadata.

No external reranker integrations are included in core. Future providers can implement RerankerProvider.

from mneno import MemoryClient
from mneno.providers.reranker import DummyRerankerProvider

client = MemoryClient(
    reranker_provider=DummyRerankerProvider(),
)

client.add("User is building Mneno, an AI memory SDK.", importance=0.5)
client.add("Unrelated high-priority note.", importance=1.0)

results = client.search(
    "AI memory SDK",
    use_reranker=True,
)

for result in results:
    print(result.memory.content)
    print(result.original_rank)
    print(result.reranked_rank)
    print(result.rerank_reason)

Use use_reranker=False to skip reranking even when a provider is configured. If use_reranker=True is passed without a provider, Mneno raises a clear provider error.

Memory Extraction

Mneno can extract structured memories from raw text. The default extractor is deterministic and local: it uses simple heuristics to identify durable facts, preferences, goals, constraints, and project information.

from mneno import MemoryClient

client = MemoryClient()

result = client.extract_memories(
    "The user is building Mneno. They prefer Python 3.11."
)

for memory in result.extracted:
    print(memory.content, memory.memory_type, memory.reason)

client.add_from_text(
    "The user is building Mneno. They prefer Python 3.11."
)

LLM-Assisted Extraction

LLM usage is optional. Mneno core does not include real provider dependencies; future adapters can implement LLMProvider to connect any model. Deterministic extraction remains the fallback.

from mneno import MemoryClient
from mneno.providers.llm import DummyLLMProvider

client = MemoryClient(llm_provider=DummyLLMProvider())

result = client.extract_memories(
    "The user is building Mneno. They prefer Python 3.11.",
    use_llm=True,
)

LLM-assisted compaction is intentionally limited: deterministic policy still decides what to keep, merge, or discard. The LLM may only improve merged memory wording for groups already selected by the deterministic engine.

Next Milestone

Mneno Core 0.3.0a1 includes the local runtime, persistence, providers, lifecycle management, sessions, tracing, and evaluation infrastructure. The next milestone is the separate Mneno Bench distribution with LOCOMO, LongMemEval, and BEAM adapters. CLI, MCP, and editor integrations remain distribution layers on top of the core SDK.

License

Apache-2.0