Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/vectorize-io/hindsight/llms.txt

Use this file to discover all available pages before exploring further.

A memory bank is the top-level container for everything Hindsight knows in a given context: facts, entities, knowledge graph connections, observations, directives, and configuration. Banks are fully isolated — memories stored in one bank are invisible to all others. You can run dozens of banks in a single Hindsight deployment, each with its own identity, extraction rules, and reasoning behavior.

​
Creating a bank

You do not need to pre-create a bank. Hindsight creates it automatically with default settings on first use. When you want explicit control over mission, disposition, or directives, create the bank before your first retain() call. 
from hindsight_client import Hindsight

client = Hindsight(base_url="http://localhost:8888")

# Minimal bank — uses all defaults
client.create_bank(bank_id="support-agent")

# Bank with identity and disposition
client.create_bank(
    bank_id="support-agent",
    name="Customer Support Assistant",
    description="Handles product support questions with empathy and clarity.",
)

​
Mission

The reflect_mission is a first-person narrative that gives the reflect() agent its identity, framing, and reasoning context. It is injected at the start of every reflect() call. 
e.g. You are a senior engineering assistant for Acme Corp.
     Always ground answers in documented decisions and rationale.
     Be direct and precise. Ignore speculation.
Set the mission via the bank config API: 
client.update_bank_config(
    bank_id="support-agent",
    config={
        "reflect_mission": (
            "You are a customer support assistant for Acme Corp. "
            "Respond with empathy and clarity. "
            "Always acknowledge the customer's concern before offering a solution."
        ),
    },
)
reflect_mission only affects reflect(). To steer what gets extracted during retain(), use retain_mission. To control what gets synthesized into observations, use observations_mission. Each operation has its own mission setting.

​
Disposition traits

Disposition shapes how the reflect() agent interprets information and frames its response. Three numeric traits on a 1–5 scale control the agent’s personality.
TraitScaleLow (1)Default (3)High (5)
Skepticism1–5Trusting — accepts information at face valueBalancedSkeptical — questions and doubts claims
Literalism1–5Flexible — reads between the linesBalancedLiteral — takes things exactly as stated
Empathy1–5Detached — focuses on facts and logicBalancedEmpathetic — considers emotional context
# Customer support — trusting, flexible, highly empathetic
client.update_bank_config(
    bank_id="support-agent",
    config={
        "disposition_skepticism": 2,
        "disposition_literalism": 2,
        "disposition_empathy": 5,
    },
)

# Code review assistant — skeptical, precise, direct
client.update_bank_config(
    bank_id="code-review",
    config={
        "disposition_skepticism": 4,
        "disposition_literalism": 5,
        "disposition_empathy": 2,
    },
)

# Legal analysis — highly skeptical, exact interpretation
client.update_bank_config(
    bank_id="legal-assistant",
    config={
        "disposition_skepticism": 5,
        "disposition_literalism": 5,
        "disposition_empathy": 2,
    },
)
Disposition only affects reflect(). It has no effect on retain() or recall().

​
Directives

Directives are hard rules injected into every reflect() prompt. Where disposition shapes tone and interpretation, directives are enforced strictly — responses that violate them are rejected. Use directives for constraints that must never be violated:
  • Compliance rules: “Never recommend specific stocks or financial products”
  • Privacy constraints: “Never share personal data with third parties”
  • Style requirements: “Always respond in formal English”
  • Domain guardrails: “Always cite sources when making factual claims”
# Add a compliance directive
client.create_directive(
    bank_id="financial-advisor",
    name="No specific stock recommendations",
    rules=["Never recommend specific stocks or financial products by name."],
)

# Add a style directive
client.create_directive(
    bank_id="support-agent",
    name="Formal language",
    rules=["Always respond in formal English.", "Never use contractions."],
)

# List active directives
directives = client.list_directives(bank_id="support-agent")
for d in directives:
    print(d.name, d.rules)

# Remove a directive
client.delete_directive(bank_id="support-agent", directive_id="dir-001")
Directives only affect reflect(). They are not applied during retain() or recall().

​
Reading and updating configuration

Bank configuration is managed separately from bank identity. This lets you change extraction rules, observation settings, and disposition independently without recreating the bank. 
# Read current config (resolved defaults + overrides)
config = client.get_bank_config(bank_id="support-agent")
print(config.config)       # fully resolved configuration
print(config.overrides)    # only fields explicitly overridden for this bank

# Update a subset of config fields
client.update_bank_config(
    bank_id="support-agent",
    config={
        "retain_mission": "Focus on product feedback and feature requests. Ignore greetings.",
        "retain_extraction_mode": "verbose",
        "observations_mission": "Observations are stable patterns in customer feedback and recurring issues.",
    },
)

# Reset all overrides — revert to server defaults
client.reset_bank_config(bank_id="support-agent")

​
Configuration reference

FieldAffectsDescription
reflect_missionreflect()First-person identity and reasoning context for the agent
disposition_skepticismreflect()How skeptical the agent is (1–5)
disposition_literalismreflect()How literally the agent interprets content (1–5)
disposition_empathyreflect()How much the agent weights emotional context (1–5)
retain_missionretain()Steers which facts are extracted
retain_extraction_moderetain()concise, verbose, or custom
retain_custom_instructionsretain()Replaces built-in extraction rules when mode is custom
retain_chunk_sizeretain()Characters per chunk (default 3,000)
observations_missionconsolidationControls what patterns become observations
enable_observationsconsolidationToggle automatic consolidation on or off

​
Bank templates

For common use cases, you can apply a preset configuration that sets mission, disposition, and extraction mode together. Presets are a starting point — you can override any field after applying one. 
# Apply a preset, then customize
client.create_bank(bank_id="support-agent", template="customer-support")

# Override a specific field after applying the template
client.update_bank_config(
    bank_id="support-agent",
    config={"disposition_empathy": 4},
)

Build docs developers (and LLMs) love

Get started for freeTalk to us