MemoSift

Seven layers.
One pipeline.

Every piece of content flows through a dual-path pipeline — a fast sync path that returns in under a millisecond, and an async path that extracts durable knowledge in the background.

SYNC<1ms
ASYNC~3s
TYPES23
MODES3
FRAMEWORKS4
SCROLL TO SEE THE PIPELINE

Three integration modes.

One SDK, three ways to use it. Pick the mode that matches what your agent needs — start with the lightest, graduate when you need more.

ModeNetworkMutates tool resultsProvides recallProvides compression
A — InspectorNoneNoNoNo
B — Sidecar (default)YesNoYes (cross-session)No
C — Co-pilotYesYes (≥2 KB stubbed)YesYes (in-session)

Mode A

Inspector

Run the local primitives — classify, extract_metadata, scan — directly. No cloud, no API key, no network. The Inspector primitives are MIT-licensed and free forever.

Use when

  • ·Compliance scanning before tool results hit the LLM
  • ·Telemetry on tool-call shapes and content types
  • ·Routing decisions based on detected content type
  • ·Local secret/PII redaction with FLAG / REDACT / BLOCK modes

Mode B

Sidecar

Push every turn to the MemoSift cloud and recall from a queryable knowledge graph. Tool results are observed but not mutated — your agent's runtime behavior is unchanged. Cross-session memory survives restarts.

Use when

  • ·Long-running agents that need persistent memory
  • ·Multi-turn workflows where a later turn benefits from facts from an earlier session
  • ·Audit trails and replayable session snapshots
  • ·Knowledge-graph queries: 'What did this user mention about X?'

Mode C

Co-pilot

Adds context-budget-driven compression on top of Sidecar. Tool results above ~2 KB are silently stubbed with artifact references in the model's view; the originals are stored and recallable. Old turns auto-compress when the context window fills.

Use when

  • ·Agents that hit context-window limits during long sessions
  • ·Pipelines that fetch large CSV / JSON / log payloads frequently
  • ·Frameworks (Claude Agent SDK, OpenAI Agents) that benefit from automatic interception
  • ·Production deployments that want auto-everything with no manual budget tuning

One call.
Two timelines.

MemoSift splits every tool result into a fast sync path that returns before the agent notices, and a slower async path that turns bytes into durable knowledge.

FAST LANE · SYNC · <1msDEEP LANE · ASYNC · ~3sAGENTtool.run()SYNC · <1msPIPELINEintercept · scan · compressASYNC · ~3sEXTRACTORmemories · entities · auditLLMcompact handleSTOREartifacts · graph
SYNCreturns before the agent moves to the next line
ASYNCextracts memories, entities, and audits in the background
RECALLbrings durable knowledge back into any later turn
THE PIPELINE
01intercept
02scan
03extract
04intent
05recall
06compress
07audit
0% COMPLETE
SYNC · <1ms
01

Intercept & Externalize

Small results pass through. Heavy results get stored.

500 CHAR GATE
log_line180 chars→ to LLM
<500 · pass through ≥500 · externalize
< 500 chars → flows to LLM untouched
≥ 500 chars → classified, stored, replaced with typed handle
23 content types · 14 languages · SHA-256 dedup
THE PIPELINE
SYNC · <1ms
01

Intercept & Externalize

Small results pass through. Heavy results get stored.

500 CHAR GATE
log_line180 chars→ to LLM
<500 · pass through ≥500 · externalize
< 500 chars → flows to LLM untouched
≥ 500 chars → classified, stored, replaced with typed handle
23 content types · 14 languages · SHA-256 dedup
SYNC · <1ms
02

Scan & Protect

Caught before storage, not after.

payload.jsonscanning…
SECRETS
0/25+
PII
0/6 types
INJECTION
0/8+
25+ secret patterns · 6 PII categories · 8+ injection vectors
Client-side — nothing sensitive reaches the wire
Three modes: flag · redact · block
ASYNC · ~3s
03

Extract Memories

Facts and entities pulled from the content itself.

RAW TURN

Q3 revenue reached $42M across EMEA and APAC, with 18% YoY growth in Germany. The EMEA team onboarded 12 new customers in September, led by SAP and Siemens.

◆ MEMORIES
◎ ENTITIES
EMEAregionAPACregion
Typed memories with importance scores and evidence quotes
Entities → project-scoped registry with fuzzy merge
Typed relations: located_in · part_of · covers · produces
ASYNC · per-turn
04

Track Intent

What the agent is doing — and was doing.

v1turn 1● ACTIVE
Explore the codebase and dependencies
RECALL CAN ASK
ms.recall(intent_version=2)
retrieve memories from a past goal epoch
Session intent versioned on every turn
Goal shifts tracked across the session lifecycle
Recall and compression organized by intent epoch
SYNC · ~750ms p50
05

Recall on Demand

Three modes. One API. Context reconstructed.

>_ ms.recall("auth middleware")
VECTOR
0 hitsscore
KEYWORD
0 hitsscore
GRAPH
0 hitsscore
CROSS-ENCODER RERANK
0.95
0.87
0.82
0.74
0.61
TOP MATCH · 98.9% CONFIDENCE
“The auth middleware validates JWT tokens against the refresh endpoint and enforces the role claim — implemented at turn 15, intent v3.”
Query — vector + BM25 keyword + entity graph
Intent — recall by goal epoch
Temporal — snapshot at any turn number
SYNC · <50ms
06

Compress Context

Old turns become knowledge, not vague summaries.

BEFORE · 146k tokens
reasoningLet me think about this...4.2k
tool_result[48 KB CSV of customer data]48.1k
reasoningThe customer data shows...3.1k
tool_result[91 KB PDF extract]91.4k
messageCustomer Q3 summary:0.1k
STAGE 1SDK · strip + replace
STAGE 2Cloud · knowledge template
AFTER · tokens
KNOWLEDGE BLOCK · intent v3
goalWire OAuth refresh tokens
memories3 high-importance · 5 medium
entitiesAPIServer · UserModel · OAuthClient
artifacts2 refs (fetch on demand)
Stage 1 (SDK) — strip reasoning, replace large results
Stage 2 (Cloud) — knowledge template by intent version
Pure SQL + template · no LLM paraphrase
ASYNC · per-turn
07

Audit & Comply

HIPAA · PCI DSS · SOX · GDPR

HIPAA0 events
PHI · identifiers
risk
PCI DSS0 events
Cardholder data
risk
SOX0 events
Financial controls
risk
GDPR0 events
Personal data
risk
SESSION DIGEST · 22 turns
overall_riskmoderate (0.28)
flagged_turns3 · [7, 14, 19]
exec_summary“Session handled PHI in turn 7 (redacted) and PII in turns 14 + 19 (flagged)…”
Every turn audited per enabled framework
Session digests track risk trajectory
Project reports with executive summaries

Seven layers.
One SDK call.

Wrap your agent loop once. MemoSift runs the pipeline on every tool result — your code doesn't change.

agent.pypython
from memosift import MemoSift

ms = MemoSift(project="acme-oauth")

# wrap the whole agent loop once
with ms.session() as s:
    for turn in agent.run():
        result = turn.tool_result
        # ↓ 7 layers run here automatically
        handle = s.process(result)
        agent.respond(handle)

    # at any point, ask for context
    ctx = s.recall("auth middleware")
01interceptsync
02scansync
03extractasync
04intentasync
05recallsync
06compresssync
07auditasync